This study employed a pooled cross-sectional survey design targeting workers in the South Korean forest industry. Stratified sampling was applied based on the six major sectors defined by the Korean Forest Industry Special Classification: (1) Input Industry, (2) Forest Product Production, (3) Processing and Manufacturing, (4) Wholesale, Retail, and Transportation, (5) Recreation and Service, and (6) Support Services.

Data were collected via face-to-face interviews using a structured questionnaire. The survey was administered in three waves: July 2022 (n = 654), August–September 2023 (n = 500), and May–June 2024 (n = 668). Participants were independently recruited in each wave (i.e., not a longitudinal panel). Given the target population and sampling procedures were consistent across the three waves, the data were pooled for analysis. The final analytic sample comprised 1,822 respondents after excluding incomplete questionnaires. Participant characteristics and the distribution across the six industry sectors are presented in Table 1.

The survey instrument assessed perceived safety risk (PSR), job satisfaction (JS), job stress (JStr), work engagement (WE), and turnover intention (TI). Items were administered in Korean. For scales originally developed in English, a translation and expert review process was conducted to support conceptual and linguistic equivalence. The composition of each measure and its sources is summarized in Table 2.

PSR was assessed using a single item developed for this study. Respondents rated the perceived likelihood of industrial accidents in their current job on a five-point scale.

JS was measured with 12 items informed by Herzberg’s two-factor theory and the Job Characteristics Model (Herzberg, 1959; Hackman and Oldham, 1976), covering intrinsic and extrinsic aspects of satisfaction. JStr was assessed using the short form of the Korean Occupational Stress Scale (KOSS), comprising 24 items across seven domains (Chang and Koh, 2005). Among these domains, the “authoritarianism and inequality” domain captures organizational-culture stressors (e.g., hierarchical climate, inconsistent instruction, perceived unfairness), alongside domains capturing workload-related stressors. WE was assessed using 12 items adapted from prior engagement literature (Kahn, 1990; Schaufeli et al., 2002). The items were intended to capture motivational facets of engagement, and their dimensionality was examined empirically using the item-screening and EFA procedures described in Section 2.4. TI was assessed using four items adapted from turnover models and prior studies in the Korean forestry workforce (Mobley et al., 1979; Kim et al., 2009); in SEM analyses, TI was specified as a latent construct indicated by these four items.

Across all multi-item scales, higher scores indicate higher levels of the construct; negatively worded items were reverse-coded prior to the 0–100 standardization described in Section 2.3.

Likert-type items were linearly transformed to a 0–100 scale. Scores were standardized using the formula (x − 1)/(k −1) × 100, where (x) denotes the observed response and (k) denotes the maximum response category (4 or 5). Negatively keyed items in the multi-item scales were reverse-coded prior to transformation. This linear transformation was applied to facilitate interpretability across measures with different Likert ranges; as a monotonic rescaling, it preserves relative ordering and does not affect inferences based on standardized relationships. Data were screened for completeness and consistency prior to analysis. Multivariate normality was examined using the critical ratio of multivariate kurtosis reported by AMOS. Bootstrap procedures were used for indirect-effect inference, as described in Section 2.4 (Kline, 2023).

Analyses were conducted using IBM SPSS Statistics 26.0 and AMOS 26.0. The analytic workflow included group comparisons, measurement evaluation, and structural equation modeling. First, reliability and exploratory factor analysis (EFA) were conducted for the multi-item scales (JS, JStr, WE, and TI). Internal consistency was evaluated using Cronbach’s α. Item screening was performed using corrected item-total correlation (CITC) and Cronbach’s α if item deleted (CAID); items with corrected item–total correlations below 0.30 were removed (Norušis, 1993). EFA was conducted using principal component analysis (PCA) with Varimax rotation (Kaiser, 1958; Abdi and Williams, 2010). Sampling adequacy and factorability were assessed using the Kaiser–Meyer–Olkin (KMO) measure and Bartlett’s test of sphericity (Bartlett, 1951; Kaiser, 1974). Factors were retained based on eigenvalues greater than 1.0, and items with factor loadings of at least 0.40 were retained (Stevens, 2002). Multicollinearity was screened using Pearson correlation coefficients and variance inflation factors (VIFs).

Group differences across industry sectors and demographic characteristics were examined using independent-samples t-tests and one-way ANOVA (Field, 2024). Homogeneity of variance was assessed using Levene’s test (Levene, 1960). When variances were unequal, Welch’s t-test or Welch’s ANOVA was applied, followed by Games–Howell post-hoc comparisons; otherwise, Scheffé’s test was used for post-hoc comparisons (Welch, 1951; Scheffé, 1953; Games and Howell, 1976; Delacre et al., 2017).

Second, SEM was performed using the two-step approach (Anderson and Gerbing, 1988). A confirmatory factor analysis (CFA) was used to evaluate the measurement model and to assess convergent and discriminant validity. The structural model was then estimated to test hypothesized paths among PSR, JS, JStr, WE, and TI. Model fit was evaluated using the Chi-square ( χ 2 ) statistic, Comparative Fit Index (CFI), Tucker–Lewis Index (TLI), and Root Mean Square Error of Approximation (RMSEA), and the Standardized Root Mean square Residual (SRMR), interpreted using commonly applied guidelines (Tucker and Lewis, 1973; Bentler, 1990; Browne and Cudeck, 1992; Hu and Bentler, 1999). Indirect effects were evaluated using bias-corrected bootstrapping with 5,000 resamples and 95% confidence intervals (Bollen and Stine, 1992). This procedure was implemented using the phantom-variable approach in AMOS to obtain estimates for specific indirect pathways. Finally, as a stability check, the CFA measurement model was re-estimated in two stratified random split-half subsamples by industry sector.

Consistent with the analysis plan (Section 2.4), psychometric evaluation focused on the multi-item scales (JS, JStr, WE, and TI); PSR, assessed with a single item, was not included in factor-analytic procedures. Item analysis was conducted to refine the measurement scales using pre-specified item-screening criteria (CITC < 0.30 and alpha-if-item-deleted diagnostics). Six items were removed from the job stress scale (Supplementary Table S1), whereas all items for JS, WE, and TI were retained. The removed items included content related to creativity and skill demands, decision latitude, future uncertainty, fairness of personnel systems, and dining-culture discomfort. These aspects may be less salient or more heterogeneously interpreted in this workforce than in the general working populations for which the original instrument was developed. Potential implications for domain coverage are addressed in the limitations. Subsequently, EFA using PCA with Varimax rotation was performed on the refined items. The KMO measures and Bartlett’s tests supported the suitability of the data for factor analysis (see Supplementary Table S2 for details). The analysis identified distinct factor structures corresponding to specific dimensions: (1) JS (4 factors): Job & Performance, Reward & Support, Relationship & Recognition, and Autonomy & Balance; (2) JStr (4 factors): Lack of Growth & Autonomy, Lack of Support & Rest, Authoritarianism & Inequality, and Work Overload; (3) WE (2 factors): Job Identity and Psychological Empowerment; (4) TI (1 factor): Single-factor structure.

All retained items exhibited factor loadings greater than 0.40 (see Supplementary Table S2 for the full factor loading matrix). Internal consistency coefficients (Cronbach’s α) for the final constructs were 0.907 (JS), 0.877 (JStr), 0.893 (WE), and 0.844 (TI).

Descriptive statistics and bivariate associations among the study variables are presented in Table 3. All variables were standardized to a 0–100 scale (PSR: higher scores indicate higher perceived accident likelihood). In the bivariate correlations, turnover intention (TI) was negatively associated with job satisfaction (JS) and work engagement (WE) and positively associated with job stress (JStr). PSR was positively associated with TI and JStr and negatively associated with JS and WE; ranges by sub-dimensions are reported in Table 3, with the full sub-factor correlation matrix provided in Supplementary Table S3. Given the pooled cross-sectional design, these coefficients are interpreted as descriptive associations and do not support causal inferences.

Sectoral differences across the six industry sectors were examined using Welch’s ANOVA with Games–Howell post-hoc comparisons (Table 4). Significant differences were observed across sectors for PSR, JS, JStr, WE, and TI. Post-hoc comparisons indicated that the Forest Product Production sector reported higher PSR alongside higher JStr and TI, and lower JS and WE, relative to several other sectors (Table 4).

Group differences by demographic and job characteristics are summarized in Table 5, with additional comparisons reported in Supplementary Table S4. PSR varied significantly across age and tenure groups, whereas WE and TI did not differ by age. Differences by tenure were observed across PSR, JS, JStr, WE, and TI (Table 5). By employment type, regular employees reported higher JS and WE and lower TI than non-regular employees, while PSR and JStr did not differ materially between these groups. Workplace type also showed differences in PSR, JS, and TI: field-site workers reported higher PSR and TI, whereas office workers reported higher JS (Table 5).

Following the two-step approach (Section 2.4), CFA was conducted to evaluate the measurement model for the four latent constructs (JS, JStr, WE, and TI) using the refined indicators retained after the pooled-sample item screening and EFA procedures; PSR (0–100; higher scores indicate higher perceived accident likelihood) was treated as an observed single-item variable and was allowed to correlate with the latent constructs. The model demonstrated an acceptable fit to the data ( χ 2 = 932.081, df = 81, CFI = 0.940, TLI = 0.923, RMSEA = 0.076, SRMR = 0.044). All standardized factor loadings were statistically significant (p < 0.001) and ranged from 0.513 to 0.890 (JS: 0.658–0.879; JStr: 0.513–0.866; WE: 0.778–0.890; TI: 0.722–0.778), supporting convergent validity. To assess the stability of the measurement structure, the same CFA specification (based on the pooled-sample refined indicators) was re-estimated in two stratified random split-half subsamples by industry sector; fit indices were comparable across the two subsamples and the pooled model (Supplementary Table S5). In the subsequent structural model, PSR was specified as an exogenous observed predictor with structural paths to the psychosocial constructs (and TI), rather than being modeled only as a correlated observed variable.

Based on the validated measurement model, the structural relationships among perceived safety risk (PSR), job satisfaction (JS), job stress (JStr), work engagement (WE), and turnover intention (TI) were estimated (Table 6 and Figure 1). Model fit indices and standardized path coefficients are reported in Table 6.

Following the ordering of paths in Table 6, PSR was negatively associated with JS and positively associated with JStr. PSR also showed a positive direct association with TI, whereas its direct association with WE was not statistically significant.

Next, JS was negatively associated with JStr and positively associated with WE. The direct path from JS to TI did not show a statistically significant association.

Finally, JStr was negatively associated with WE and positively associated with TI, and WE was negatively associated with TI (Table 6). As a sensitivity check, we compared the proposed model with two nested alternative specifications (PSR → WE constrained to zero; PSR → TI constrained to zero), and comparative fit statistics are reported in Supplementary Table S6.

To estimate indirect effects, bias-corrected bootstrapping with 5,000 resamples was conducted using the phantom-variable approach. The specified indirect pathways are illustrated in Figure 2 (PSR pathways) and Figure 3 (JS pathways), and Table 7 reports the estimated indirect effects with bias-corrected 95% confidence intervals.

For perceived safety risk (PSR), the total indirect effect on turnover intention (TI) was statistically significant (i.e., the 95% bias-corrected confidence interval excluded zero). Among the specific indirect effects, significant pathways were observed through job satisfaction (JS) and job stress (JStr), including serial pathways involving work engagement (WE), as well as pathways operating via JStr.

For job satisfaction (JS), the total indirect effect on TI was also statistically significant (i.e., the 95% bias-corrected confidence interval excluded zero). Significant specific indirect effects were observed via JStr, via WE, and via the serial pathway through JStr and WE.

Additional indirect pathways were estimated; however, pathways whose 95% bias-corrected confidence intervals including zero are omitted from Table 7 for clarity.

Using a five-variable model, this study examined perceived safety risk (PSR), job satisfaction (JS), job stress (JStr), work engagement (WE), and turnover intention (TI) within a single structural framework. As summarized in Table 6, PSR was linked to less favorable psychosocial profiles (lower JS and higher JStr) and showed a positive association with TI, while the direct PSR–WE path was not supported.

Among the psychosocial variables, the pattern was consistent with the JD–R framing: stress- and engagement-related processes were most closely aligned with TI, and JS related to TI primarily via its associations with lower stress and higher engagement (Demerouti et al., 2001).

The mediation analysis further clarified how PSR and JS relate to TI through indirect pathways (Table 7 and Figures 2, 3). For PSR, the total indirect effect on TI was statistically significant (i.e., the 95% bias-corrected confidence interval excluded zero), with supported specific pathways operating through JS and JStr, including serial configurations involving WE. For JS, the total indirect effect on TI was also statistically significant, with supported specific pathways via JStr, via WE, and via the serial pathway through JStr and WE (Table 7). This configuration is consistent with mediation typologies that distinguish indirect-only patterns from direct effects in models with multiple mediators (Zhao et al., 2010). Other estimated indirect pathways whose bias-corrected confidence intervals included zero are omitted from Table 7 for clarity.

Bivariate associations and sub-dimension correlations provide descriptive context for interpreting the model (Table 3; Supplementary Table S3). The sub-dimension patterns suggest that relational/structural stressors (e.g., authoritarian climate and perceived inequality) may be more closely aligned with TI than workload-related indicators, while identification-related aspects of WE may be salient correlates of TI. In addition, group comparisons indicated heterogeneity in PSR and psychosocial variables across sectors and workplace settings (Tables 4, 5). The Forest Product Production sector and field-site workers showed profiles characterized by higher PSR and less favorable psychosocial outcomes relative to several comparison groups (Tables 4, 5). These descriptive patterns are consistent with the interpretation that safety appraisals and psychosocial conditions co-occur in specific production- and field-oriented contexts.

The results suggest several practical considerations for workforce sustainability. First, safety management may be relevant not only for accident prevention but also for turnover-related attitudes insofar as PSR is associated with psychosocial variables and TI within the model (Tables 6, 7). Second, the descriptive sub-dimension patterns indicate that organizational fairness, communication, and supervisory practices may complement workload-oriented approaches (Supplementary Table S3). Third, strengthening job identity (e.g., role clarity, recognition, and professional meaning) may be relevant for engagement-related components inversely associated with TI. From a forest management perspective, stabilizing the workforce is not merely a human resource concern but a strategic consideration for maintaining operational continuity and implementing long-term management plans effectively.

Several limitations should be considered. First, the cross-sectional design does not support causal inference; the directional paths represent theory-consistent associations. Accordingly, reverse causality (e.g., higher strain shaping satisfaction) and omitted variable bias (e.g., unmeasured organizational or personal factors) cannot be ruled out, and the estimated paths should be interpreted as theory-consistent associations rather than causal effects. Second, measures were self-reported, which may introduce common method variance. Third, PSR was assessed using a single item; future studies could employ multi-item measures capturing safety climate and related dimensions. Fourth, item screening for the job stress scale improved internal consistency in this sample but may have narrowed content coverage relative to the full KOSS domains; future work should confirm the retained structure in an independent sample and evaluate domain coverage with alternative stress measures. Finally, psychometric refinement was conducted using pooled-sample item screening and EFA for established scales, followed by CFA; to address concerns about capitalization on chance, a stratified split-half stability check was conducted and yielded comparable fit indices across subsamples (Supplementary Table S5). Additionally, because the dataset is not publicly available due to institutional regulations, independent reproducibility may be constrained; access is subject to the process described in the Data availability statement. Future studies could combine survey data with administrative indicators and multi-item safety measures to evaluate whether perceived risk aligns with objective safety conditions. Longitudinal or repeated-measures designs and independent-sample validation would further help assess temporal ordering and the stability of the measurement and structural relations across survey waves and subpopulations.