The organization’s Employees Survey is an annual and anonymous employee survey of work attitudes and experiences. Sixteen items from the Employees Survey were completed at T0, and 17 items were completed at T1. Previous exploratory factor analysis (EFA) with a subsample of respondents at T0 (Doyle et al., 2016) indicated that the items reflected two dimensions: “job satisfaction” (five items) and “instrumental support” (five items). Notably, instrumental support refers to more tangible help or information such as assistance with solving a problem or with performing a difficult task (Semmer et al., 2008; Hergatt Huffman and Frevert, 2013; Peeters et al., 2014). For the present study, the Instrumental Support Scale comprised items assessing, for example, the extent to which respondents felt well-informed about what was expected in their job; had the necessary tools and equipment to perform their job; and experienced cooperation from colleagues when performing jobs.

All items of the Job Satisfaction Scale (JS Scale) and Instrumental Support Scale (IS Scale) were measured on a five-point Likert scale, including scales ranging from “very satisfied” to “very dissatisfied,” “strongly agree” to “strongly disagree,” and “very good” to “very poor.” Total scores for job satisfaction and instrumental support were computed by averaging scores ranging from 1 to 5. In both cases, higher scores reflect higher levels of each variable. The reliability estimates for both scales were satisfactory: Cronbach’s alpha = 0.79 at T0 and 0.80 at T1 (job satisfaction), and 0.74 at T0 and 0.76 at T1 (instrumental support). Given that Doyle et al. (2016) used EFA, the validity of the factor structure of the Employee Survey items was tested in the present study using confirmatory factor analysis (CFA).

Resilience was measured using the Dispositional Resilience Scale-15 (DRS-15) (Bartone, 1999, 2007). The decision to use the DRS-15 was based on its established validity, acceptable internal consistency, and acceptable test–retest reliability, in addition to its brevity (Bartone, 1995, 2007). The DRS-15 uses both positively and negatively keyed items, and includes three factors of resilience: commitment, control and challenge (Bartone, 2006), each measured by five items scored on a four-point scale ranging from “not at all true” to “completely true.” Example items comprise: “Most of my life gets spent doing things that are meaningful” (commitment), “By working hard, you can nearly always achieve your goals” (control), and “Changes in routine are interesting to me” (challenge) (Bartone et al., 2012; Kelly et al., 2014). When the six negatively keyed items are reversed, a total score for resilience can be calculated by adding scores for all items (Hystad, 2012). For the present study, as numerous respondents were missing scores for particular items of the DRS-15, a total score for each respondent was calculated by averaging rather than adding scores.

Research by Bartone (1995) reports internal consistency for the total scale (α = 0.83) and three subscales of commitment, control and challenge (α ranging from 0.70 to 0.77) that equal or exceed the acceptable alpha threshold of 0.70 (see Tavakol and Dennick, 2011). In another study conducted by Bartone (2007), the 3-week test–retest reliability coefficient for the DRS-15 was 0.78, exceeding the recommended threshold of above 0.70 (Kline, 2000). However, the test–retest coefficients for the three subscales were 0.75 for Commitment, 0.58 for Control, and 0.81 for Challenge, indicating a test–retest coefficient for Control that was below the recommended threshold (Bartone, 2007). While Doyle et al. (2016) reported the internal consistency for the total DRS-15 score as 0.72, the internal consistency was 0.65 for Commitment, 0.57 for Control, and 0.57 for Challenge, which were below the acceptable alpha value of 0.70. Accordingly, total resilience scores were used in the present study. The internal consistency for the DRS-15 was 0.70 at TO and 0.73 at T1.

Perceived stress was measured using the Perceived Stress Scale-4 (PSS-4). The PSS-4 is a four-item version of the Perceived Stress Scale developed by Cohen and colleagues, which measures an individual’s assessment of stressful situations in the last month (Cohen et al., 1983). The decision to use the PSS-4 was based on its validity, acceptable internal consistency, and brevity (Cohen et al., 1983; Cohen and Williamson, 1988; Warttig et al., 2013). The PSS-4 comprises two positively stated and two negatively stated items, with a response set ranging from 0 (never) to 4 (very often) (Wu and Amtmann, 2013). An example item is: “In the last month, how often have you felt that you were unable to control the important things in your life?” Positively stated items are reverse coded prior to summing the items, and higher scores indicate higher perceived stress (Wu and Amtmann, 2013). For the present study, as numerous respondents were missing scores for particular items of the PSS-4, total scores were calculated using average rather than summed scores.

Cohen et al. (1983) reported the internal consistency for the PSS-4 as 0.72, exceeding the acceptable alpha threshold of 0.70 (see Tavakol and Dennick, 2011). In the same study, Cohen et al. (1983) reported the test–retest reliability over a 2-month interval as 0.55, below the recommended threshold of above 0.70 (Kline, 2000). In another study comprising a probability sample of the United States (N = 2,387), Cohen and Williamson (1988) reported the internal reliability for the PSS-4 (α = 0.60) as less than that of the 10-item version (α = 0.78) and 14-item version (α = 0.75). While the PSS-4 indicates a moderate loss in internal reliability relative to the 14-item scale, the brevity of this scale is advantageous when time for assessment is limited (Warttig et al., 2013). For the present study, the internal consistency for the PSS-4 was 0.55 at both T0 and T1.

The study’s theoretical model of perceived stress and job satisfaction amongst merchant seafarers was tested using structural equation modeling (SEM). SEM incorporates two analytical procedures: CFA, which evaluates the measurement component of a theoretical model, and path analysis, which evaluates the relationship between latent variables (Hyland et al., 2015). The “measurement model” is therefore synonymous with CFA by specifying the way in which each measure loads onto a certain factor, while the “structural model” specifies the way in which certain latent variables directly or indirectly influence changes in the values of other latent variables in the model (Byrne, 2006). Notably, an advantage of SEM is that it controls for measurement error in latent outcomes (Leite, 2017).

SEM entails an assessment of how well a proposed theoretical model corresponds with covariance data obtained from a sample (Sivo et al., 2006). The adequacy of a model is determined in relation to a number of “model fit” indices, and standard recommendations (Hu and Bentler, 1999) indicating that a good model fit is reflected by: a chi-square to degree of freedom ratio of less than 3 to 1; Comparative Fit Index (CFI) and Tucker Lewis Index (TLI) values > 0.90; Root-Mean-Square Error of Approximation (RMSEA) and Standardized Root-Mean-Square Residual (SRMR) values < 0.08. Models were assessed using Mplus 7.0 (Muthén and Muthén, 2013) with robust maximum likelihood (MLR) estimation (Yuan and Bentler, 2000).

Dummy variables were created for the variables of job category, race/ethnicity and location. Categories with the largest frequency were omitted when generating dummy variables for job category, race/ethnicity and location, namely the categories of “Officer/Engineer,” “South Asian,” and “On Passage,” respectively. Dummy variables were formulated for job category as follows: (0 = not crew, 1 = crew), and (0 = not catering, 1 = catering). For race/ethnicity, dummy variables comprised: (0 = not Caucasian, 1 = Caucasian), (0 = not East Asian, 1 = East Asian), (0 = not Other, 1 = Other), (0 = not Latino/Hispanic, Middle Eastern, or Mixed, 1 = Latino/Hispanic, Middle Eastern, or Mixed), and (0 = not African, 1 = African). For location, dummy variables were as follows: (0 = not approaching port, 1 = approaching port), and (0 = not loading/discharging, 1 = loading/discharging).

Based on the EFA previously conducted with a subsample of T0 respondents (Doyle et al., 2016), the JS Scale and IS Scale were investigated using CFA with MLR estimation. EFA may be used as an exploratory procedure when developing a measure, followed by CFA to investigate if the structure identified during EFA works in a different sample (Harrington, 2009). Hence, while EFA was previously conducted with T0 data, CFA was conducted in the present study with T1 data. For both the JS Scale and IS Scale, the model was a one-factor solution, whereby five items loaded onto one latent variable, proposed to represent job satisfaction and instrumental support, respectively. The measurement models of both latent variables, i.e., job satisfaction and instrumental support, were then incorporated within the full structural equation model. The full structural model was tested using T0 data (N = 512). A minimum sample size of 500 has been recommended for SEM (Hazard Munro, 2005). Therefore, T0 data was used to provide a sufficiently large sample size for SEM. Accordingly, CFA was conducted with T1 data (N = 276), while the full structural equation model was tested using T0 data (N = 512).

The model produced less than satisfactory fit to the sample data as indicated by the chi-square to degrees of freedom ratio (>3:1), CFI value (0.75), TLI value (0.51), RMSEA value (0.24), and SRMR value (0.09). Inspection of the modification indices indicated an improved fit by correlating the error variances for item 2 and item 3 (MI = 90.013) (Byrne, 2012). As outlined below in Table 3, similar questions are asked by items 2 and 3, with the former asking for respondents’ level of agreement with the statement “I am proud to work for Shell,” and item 3 assessing respondents’ agreement with the statement “I would recommend Shell as a good employer.” Accordingly, the error variances of items 2 and 3 were correlated and the model produced good fit to the sample data according to the chi-square to degrees of freedom ratio (<3:1), CFI value (0.99), TLI value (0.97), RMSEA value (0.06), and SRMR value (0.01). Factor loadings for each item on the latent variable of job satisfaction were statistically significant (p < 0.001) and positive. Three items had high factor loadings (>0.60), while two items had moderately high factor loadings (>0.30) (see Table 3).

The model produced good fit to the sample data according to the chi-square to degrees of freedom ratio (<3:1), CFI value (0.97), TLI value (0.95), RMSEA value (0.05), and SRMR value (0.03).

Factor loadings for each item on the latent variable of instrumental support were statistically significant (p < 0.001) and positive. Three items had high factor loadings (>0.60), while two items had moderately high factor loadings (>0.30) (see Table 4).

The measurement models of both latent variables, i.e., the re-specified measurement model of job satisfaction and the measurement model of instrumental support, were then incorporated within the full structural model. The model produced acceptable fit to the sample data according to the chi-square to degrees of freedom ratio (<3:1), RMSEA value (0.05), and SRMR value (0.05). The CFI value (0.87) and TLI value (0.83) indicated less than satisfactory model fit. The model explained 70.6% of variance in job satisfaction scores and 23.8% of variance in perceived stress scores. Table 5 provides the regression effects produced from the SEM analyses.

As presented in Table 5, the regression analysis indicated that instrumental support significantly predicted job satisfaction, as did resilience. The ethnicities of Caucasian and African also significantly predicted job satisfaction, which is due to Caucasian and African participants reporting less job satisfaction relative to South Asians. Therefore, ethnicity significantly predicted job satisfaction. Furthermore, the job categories of crew and catering significantly predicted job satisfaction, signifying that crew and caterers reported higher job satisfaction than officers/engineers. Job category therefore significantly predicted job satisfaction.

Instrumental support significantly predicted perceived stress, as did resilience. The ethnicities of Caucasian and Other combined (Mixed, Middle Eastern, and Latino/Hispanic) also significantly predicted perceived stress, which is accounted for by these ethnicities reporting less perceived stress relative to South Asians. Ethnicity therefore significantly predicted perceived stress.

Crew and caterers reported significantly higher job satisfaction than officers/engineers. As seafarers’ ranks are conflated with ethnicity, it was anticipated that findings would instead indicate lower job satisfaction of ratings relative to officers/engineers, as the literature specifies inequities experienced by seafarers from LMICs, such as linking nationality to senior positions, longer tours of duty, and different pay for the same work (Carter, 2005; Dimitrova and Blanpain, 2010; Borovnik, 2011; Baylon and Santos, 2015; MacLachlan, 2017a). This finding was therefore unexpected, in light of literature indicating an association between organizational justice and job satisfaction (McAuliffe et al., 2009; Furnham, 2012; Aamodt, 2013; Schultz and Schultz, 2016; Bekru et al., 2017).

However, comparable with this finding, Bergheim and colleagues found no difference for job satisfaction between European and Filipino seafarers, and explained the high scores for job satisfaction of Filipino participants in relation to their more collectivist culture, which prioritizes work-group cohesion and relationships with peers, thereby explaining job satisfaction of Filipinos in terms of work-group relations (Bergheim et al., 2015). Indeed, regular interaction with others, friendships in the workplace, and emotional support may be strong predictors of job satisfaction (Morgeson and Humphrey, 2006; Meyers, 2007). Europeans rank higher on individualism than Filipinos, with the Philippines being largely a more collectivist society (Hofstede, 2001). Higher job satisfaction of ratings relative to officers/engineers in this study may therefore possibly be accounted for by greater social support amongst ratings. Support for this explanation is provided by the qualitative findings of a related study (McVeigh et al., 2018), whereby two superintendents indicated that ratings experienced a better social life on-board than officers due to ratings’ fixed meal times, their homogenous nationality, and the set-up of their mess/social area on-board, as exemplified by a superintendent’s observation: “For the crew, for the Filipino lads... they get together more as a group... The social interaction is gone at officer level.” Therefore, ratings’ higher job satisfaction levels relative to officers/engineers may possibly be explained in terms of more cohesive relationships with peers.

Dispositional resilience significantly predicted job satisfaction (β = 0.25, p < 0.001), although resilience emerged as a moderator predictor. This finding is consistent with the literature, which specifies that resilience is positively associated with job satisfaction (Youssef and Luthans, 2007; Bergheim et al., 2015; Hyde, 2015; Hyde and Knocker, n.d.). Moreover, the SEM analysis indicated that resilience significantly predicted perceived stress (β = -0.39, p < 0.001), although again it was a moderator predictor. Indeed, the SEM model explained 23.8% of variance in the criterion variable of perceived stress, and the strongest predictive effect was for dispositional resilience. This finding corresponds with previous research reporting that resilience may protect against the adverse impacts of stress (Friborg et al., 2006; Hjemdal et al., 2006; Ong et al., 2006; Pietrzak et al., 2010; Doyle et al., 2016).

Instrumental support significantly predicted job satisfaction (β = 0.72, p < 0.001), with instrumental support emerging as a strong predictor, in accordance with criteria proposed by Acock (2008). Furthermore, instrumental support significantly predicted perceived stress (β = -0.16, p = 0.005), although instrumental support was a weak predictor of perceived stress. Similarly, findings of a previous study with a subsample of respondents of the T0 work questionnaire (Doyle et al., 2016) indicated that higher instrumental support was significantly associated with lower perceived stress, and instrumental support was significantly positively associated with job satisfaction. The SEM model explained 70.6% of variance in the criterion variable of job satisfaction, explaining a considerable amount of variance for job satisfaction, and the strongest predictive effect was for instrumental work support. This finding signifies the importance of instrumental support, including more tangible help or information such as assistance with solving a problem or with performing a difficult task (Semmer et al., 2008; Hergatt Huffman and Frevert, 2013; Peeters et al., 2014). This makes intuitive sense, as with the resources to conduct work effectively, one may be more likely to experience a work environment as more satisfying and less stressful.

This study has focused on a particular organization engaged in bulk hydrocarbon transport. The specific attributes of this organization, including the distances and routes traveled, policies, practices, routines, and the multinational nature of the organization alongside its high public profile, all establish a particular working and living environment on-board that is not necessarily similar to other shipping organizations or groups of seafarers. It is critical to caution against extrapolation from one study across the maritime industry, which constitutes a wide diversity of States, employers, flags, ship types, contract types, and recruitment and remuneration practices. We have highlighted elsewhere the importance of considering context when formulating policy initiatives (McVeigh et al., 2016), and this is certainly also essential in the maritime industry.

The item in the work questionnaire assessing seafaring experience comprised overlapping response categories, namely 0–1, 1–5, 5–10, 10–20, and more than 20 years. For example, respondents with 5 years of experience may therefore have responded as having “1–5” or “5–10 years” of experience. Such overlapping response categories were not therefore mutually exclusive (McBurney and White, 2010; Grove et al., 2013).

Data was not available with respect to the number of seafarers on each vessel who were informed of the study and asked to complete the work questionnaire. It was not therefore possible to specify a response rate. Accordingly, it is possible that a sampling bias may have occurred. For example, those who completed the questionnaire may have been particularly resilient.

While the model explained a considerable amount of variation in job satisfaction scores, the model was not as effective in explaining variation in perceived stress scores. This suggests that the factors included in the model were more effective at explaining indicators of positive mental health rather than indicators of negative mental health in the study’s sample. Accordingly, for future research, this model may be more important for researchers aiming to explore indicators of positive mental health than for researchers aiming to investigate indicators of negative mental health. This is potentially one of the reasons for the slightly less than desirable CFI and TLI statistics.