Results Directory UMM :Journals:Journal of Operations Management:Vol18.Issue4.Jun2000:

4.2.5. Safety efficacy Ž . Based on Banduara’s 1986 recommendation re- garding the measurement of efficacy perceptions, we used three items that assessed safety-efficacy strength Ž . see Appendix A . In general, respondents indicated the extent to which they were confident that they could work safely and avoid safety hazards. Re- Ž . sponses ranged from 1 strongly disagree to 7 Ž . strongly agree . Coefficient a was 0.83, but for modeling purposes we used each item as a separate indicator of the safety-efficacy construct. 4.2.6. Safe work behaÕior Two items assessed safe work behavior, the crite- rion variable for our sociotechnical model. The first item asked subjects: ‘‘About what percent of the time do you follow all of the safety procedures for the jobs that you do?’’ Possible responses ranged from 0 to 100, in 10 increments. The second item was measured similarly but asked respondents to indicate the percentage of time that their cowork- ers followed safe work practices. Based on attribu- tion theory, we felt that the addition of the second question would serve as a supplemental validity check on self reports of behavior. Self-estimates of safe Ž . behavior frequency averaged 82.82 S.D.s 16.54 , ranging from 0 to 100. Estimates of coworker Ž . safe behavior averaged 72.47 S.D.s 24.40 , with a range of 0 to 100. Although the difference was Ž . significant T s 9.074; p - 0.0001 , the two items Ž . were highly correlated r s 0.51; p F 0.001 . For modeling purposes, we used the two estimates Ž . self and other as separate indicators of the safe workplace behavior construct. Because the safe work items were measured differently than the variables Ž within the other constructs percent vs. seven-point . scales , we were able to dampen some of the prob- lems associated with common method dependence. ŽAs we shall demonstrate later in this paper, we also were able to employ statistical tests to address this . possible concern.

5. Results

Table 2 summarizes the results of the plant com- parison and measurement model tests, and Table 3 provides descriptive statistics of model indicators. In reference to measurement model tests in Table 2, CFI and IFI values of 0.90 and greater indicate 8 9 10 11 12 13 14 15 16 17 18 0.55 0.36 0.31 y0.12 y0.13 y0.07 y0.14 y0.15 y0.01 0.59 y0.13 y0.16 y0.05 0.54 0.78 0.45 0.38 0.23 y0.18 y0.19 y0.15 0.65 0.53 0.31 y0.16 y0.15 y0.15 0.52 0.36 0.36 0.28 y0.16 y0.17 y0.16 0.54 0.55 y0.31 y0.31 y0.07 0.24 0.25 0.25 y0.27 y0.34 y0.21 y0.31 y0.27 y0.17 0.19 0.16 0.14 y0.21 y0.31 y0.19 0.51 Ž . adequate model fit Bollen, 1989 . PFI values of 0.60 and greater are suggested as a rule of thumb Ž criterion for model retention Williams and Pod- . sakoff, 1989 . RMSEA values of 0.08 or less indi- cate reasonable error of approximation, and values of Ž 0.05 or less indicate a close fit Browne and Cudek, . 1992 . 5.1. Tests of between-plant differences We wished to determine whether it was appropri- ate to combine data from the two plants, or whether they should be treated separately. For this purpose, we used a two-group Confirmatory Factor Analysis Ž . Ž . CFA Bollen, 1989 . We began with a model where all factor loadings were constrained to be equal Ž . Table 2, Model 2 and compared that to a model where factor loadings were free to vary across groups Ž . Table 2, Model 1 . An insignificant chi-square dif- ference indicated that the measures were interpreted similarly across plants. This justified combining data from plants A and B for further analyses. 5.2. Testing the measurement model Following determination of construct equivalence across plants, we assessed the dimensionality of the constructs in our six-construct hypothesized model Ž . Fig. 1 . Before delving into predictive relationships, we simply wanted to see if the measures made sense and whether the various constructs were theoretically independent. The results of the CFA appear in Fig. 2. Fig. 2 highlights the factor loadings of the indica- tors associated with each of the six latent constructs included in our model. Most of these results matched the expectations we had when we designed the sur- vey, although some original survey items did drop out because of their weakness in explaining latent factor variation. For example, we had identified 10 hazard categories, but the latent hazard factor was adequately measured with just five categories. As Ž shown in Table 2, this baseline model Table 2, . Model 3 fit the data well. Additionally, all standard- ized factor loadings were significant, averaging 0.72 Ž . Ž . Fig. 2 . A single factor model Table 2, Model 4 did not perform as well as the six factor model. Fig. 2. Baseline measurement model. Comparative Fit Index s 0.96; Incremental Fit Index s 0.96; Root Mean Square Error Approximations 0.05. Beyond the baseline and single factor tests for discriminant validity, we contrasted the baseline six- factor model with four other models, each of which presumed that two or more latent constructs were Ž actually measuring the same thing see Models 5, 6, . 7, and 8 in Table 2 . Sequential chi-square difference tests showed that constraining the system-level and person-level constructs to be equivalent produced significantly worse fitting models as compared with the baseline model. Moreover, decreases in CFI indi- cate a material reduction in model fit for each of the Ž . constrained models Widaman, 1985 . These results further reinforced our confidence in the indepen- dence of the constructs specified in the model. 5.3. Structural model results We began our structural analysis with a test of the system–person sequence model depicted in Fig. 1. Thus, we first examined whether system character- Ž . istics influence employees to follow or not follow safety procedures indirectly through their effects on Ž . person-based characteristics Fig. 3 . Next, we com- pared the first model to one that added direct links Ž from system characteristics to safety outcomes Fig. . 4 . Finally, the initial model was compared to a model where system and person influences on safety Ž . were separate Fig. 5 . All three of these models were tested using covariance structure analysis pro- Ž . cedures in Bentler’s 1995 EQS program. Model Ž . paths were evaluated for significance see Figs. 3–5 , and model fit was assessed by four fit indices: CFI, IFI, PFI, and RMSEA. An indirect effects model based on our hypothe- sized system–person sequence, accurately explained patterns in the sample data, as indicated by the fit indices in Fig. 3. All structural parameter estimates Ž . were significant p - 0.05 and in the predicted directions. In the direct effects model depicted in Fig. 4, we assessed the propriety of adding direct paths from system characteristics to safety outcomes. We com- pared these results with those from the system–per- Ž . son sequence indirect effects model in Fig. 3 and found that the chi-square and CFI values were not Ž 2 significantly different x difference s 3.29, 3 df ; . CFI difference s 0.00 . Moreover, the coefficients for the added model paths were insignificant. Be- cause the added paths did not allow us to explain more variance in the model, the contrast argued for accepting the more parsimonious system–person se- quence model. The next comparison assessed whether system and person influences on safety outcomes operate directly, rather than sequentially. As shown in Fig. 5, we tested this model by removing system links to personal factors, and found it to be less explanatory than the indirect system–person sequential model. The chi-square difference between the models was Ž . significant 168.64, p - 0.05 and the CFI dropped from 0.95 to 0.91. These results reinforced our argu- ments for the system–person sequence model pre- sented in Fig. 3. Of the structural models evaluated, the indirect Ž . model system–person sequence represented the best fitting parsimonious model. Results from this model Ž . Fig. 3 show that the presence of safety hazards can Ž . degrade perceived safety climate y0.32 and in- crease operators’ feelings of pressure to value expe- Ž . diency over safety 0.29 . However, the perception of a strong safety climate can diminish feelings of Ž . pressure y0.30 . Thus, safety hazards influence perceived pressure directly and indirectly through climate. The significant path coefficients support our predictions regarding system factor relationships. Fig. 3. Results: indirect effects model. The system–person sequence model for predicting safe work behaviors. Standardized path coefficients are displayed adjacent to influence arrows. x 2 s 320.29 ; Comparative Fit Index s 0.95; Incremental Fit Index s 0.95; Parsimonious Fit Index s 0.79; Root Mean Squared Error Approximations 0.06; Degrees of Freedom s 128. Indicates p - 0.05. Fig. 4. Results: dual effects model for predicting safe work behaviors. Standardized path coefficients are displayed adjacent to influence arrows. x 2 s 317.00 ; Comparative Fit Index s 0.95; Incremental Fit Index s 0.95; Parsimonious Fit Index s 0.77; Root Mean Squared Error Approximations 0.06; Degrees of Freedom s 125. Indicates p - 0.05. Structural results further show that system factors affect safe behaviors indirectly through person fac- tors. Specifically, pressure is positively related to Ž . cavalier attitude 0.87 and negatively related to Ž . safety efficacy y0.25 . Finally, person-linked fac- tors directly affect safe behaviors. A cavalier attitude Ž . results in less safe behavior y0.47 , and strong perceptions of safety efficacy lead to more safe Ž . behavior 0.23 . 5.4. Discussion Managers make many assumptions about the rea- sons why employees engage in unsafe work behav- iors. Those subscribing to the Du Pont model, espe- cially if they interpret it on a superficial level, may assume that unsafe acts emerge solely from personal characteristics of the employee, and that they have no links to contextual factors in the operating envi- ronment. Our results provide evidence for another point of view: factors internal to the organization and within the scope of operating control have a mean- ingful influence on employee safety behaviors. The solution to the problem of unsafe acts goes beyond the training approach that is often at the center of safety programs. The model we have tested and verified highlights several areas where managerial action can truly make a difference. Encouraging employees’ safety goes well beyond the slogans and posters that so often serve as surro- gates for safety programs in industrial settings. Sin- cere, concerted company efforts aimed at all five of the predictor constructs identified in our model can produce effective change with lasting effects. 5.4.1. Safety hazards Our model-testing results provide evidence that safety hazards do more than just directly cause acci- dents — they indirectly influence employees’ per- ceptions of organizational factors such as safety cli- mate and pressure, and they indirectly lead to unsafe behaviors. Although we uncovered 10 hazard cate- gories through data reduction, only five of them explained variance in our model for predicting be- havior. These included back hazards, industrial hy- giene hazards, sitting hazards, equipment unavail- ability hazards, and equipment handling hazards. Of the other five, moving object hazards proved to be the highest rated, even though they did not explain much variance. This indicates that they were seen as universally hazardous, and it suggests a useful man- agerial insight — hazards that are inherent in system Fig. 5. Results: direct effects model for predicting safe work behaviors. Standardized path coefficients are displayed adjacent to influence arrows. x 2 s 488.93; Comparative Fit Index s 0.91; Incremental Fit Index s 0.91; Parsimonious Fit Index s 0.75; Root Mean Squared Error Approximations 0.08; Degrees of Freedom s 127. Indicates p - 0.05. design may not negatively influence the safe behav- ior decision model because employees know they cannot be changed. A steel plant operator must work with molten metal in large, overhead ladles. A long- distance trucker must operate a large vehicle at high speeds. A fisher must work from a boat in high seas. This does not mean that these hazards will not cause accidents, but their omnipresence makes them less predictive of behavior. On the other hand, hazards that did explain variance were those that managers could remove or abate — employees apparently sensed that when ‘‘fixable’’ operating hazards per- sisted, managers did not really care. 5.4.2. Safety climate Although climate is often viewed as a macro-level organizational characteristic, safety climate can man- ifest a degree of variability within an organization. Some of these variabilities come from the ways in which different groups of people perceive top-level communication. For example, in the case of the steel company studied here, the department where vocal union officials happened to work gave lower safety climate ratings than some other departments where there were no vocal union officials. A positive safety climate will be characterized by an open door policy for hazard and accident reporting, a sincere concern for employee well-being, and fairness in accident investigations. If these conditions exist, our results indicate that employees are likely to work more safely. 5.4.3. Pressure Perceptions of pressure to value expediency over safety, as we have demonstrated, are influenced by hazards and safety climate. Beyond those factors, managers can also effect changes in perceived pres- sure by clearly communicating priorities and keeping safety visible during times of ramp-up. Perhaps the most well-known example in this regard may be found in the Toyota plants in Japan. Visitors to those plants often express amazement at Toyota’s commit- ment to safety. Not until safety has been established do these plants begin to focus on quality, and not until quality is under control are speed and efficiency addressed. Virtually every plant we visited was plagued by spikes in accident rates during times of increased production. In all cases, operating level employees expressed the feeling that safety programs had been suspended. They felt that if they followed safety procedures they would not be able to keep up — production bonuses, or even their jobs, would be in jeopardy. This is a problem without an easy solution, but if employees can develop methods of working safely and efficiently during times of slow produc- tion, they may be better prepared to take on heavier loads when they ramp-up. 5.4.4. CaÕalier attitudes about safety behaÕiors As we have discussed previously, people possess some innate tendencies toward risk-taking. Addition- ally, some jobs and industries will attract more than their share of people who seek or avoid risks. How- ever, organizations can influence the variables within this construct, beginning with those that fall to the Ž left of cavalier attitudes in our model i.e., hazards, . climate, pressure . In addition, they may endeavor to select employees who are not high-level risk-takers and they may introduce behaviorally-based safety programs that focus on employee work practices instead of accident outcomes. 5.4.5. Safety-efficacy As with the other constructs on the right-hand side of our model, we recommend that managers carefully consider the variables that precede safety- efficacy: hazards, safety climate, and pressure. Addi- tionally, in keeping with findings reported in the literature on self-efficacy, we prescribe hands-on training in safety procedures because going beyond classroom lectures to allow employees to test their skills at safety practices will build confidence. 5.4.6. General obserÕations The system–person sequential model holds pre- scriptions of its own: the operating environment, and its technical and social characteristics, influence em- ployee safety behaviors. Operations managers can benefit from careful consideration of the linkages we have uncovered. Additionally, they can take actions to influence the individual constructs that we have examined by addressing factors not explicitly shown in our model. Our primary recommendation is that operations managers in manufacturing settings should appreciate the interdependence among the social and technical variables in the system–person sequence model. A change along one dimension can cause a profound series of chain reactions. 5.5. Limitations Although this study adds to our understanding of safe and unsafe work behaviors, we must acknowl- edge that it presents some limitations, indicating the need for further research. First, the study was con- ducted in one firm in a single industry. This focus gives us a homogeneous sample of responses from the same environment, allowing for initial tests of instrument reliability and defensible interpretations of dimensionality. Moreover, we were able to demonstrate that our baseline measurement model was interpreted similarly in two different plants, providing some sense that the results may be repeat- able across settings. However, future research should attempt to replicate our findings in other environ- ments. A second limitation is the use of same-source data; relying on self-report data can inflate relations among modeled variables. However, we tried to Ž . reduce this potential problem by a positioning mea- sures of the indicators in different locations through- Ž . out the larger survey, b measuring the indicators Ž . with different types of response categories, c using Ž . observationally based validity checks, and d find- ing corroboration from managerial data. Further- more, we tested a same-source measurement model and found support for the multidimensional nature of our constructs. We therefore believe it is unlikely that a common method factor could have explained our results. One could also question the validity of the direc- tional arrows and sequencing in our model. For example, who is to say that pressure produces a negative safety climate, rather than the reverse? We recognized that correlation does not always indicate cause, so we tested several alternative models where sequences were altered. The net effect was that none of these models explained as much variance as the one we had built upon our preliminary field work and literature review. We would encourage more work in this area, however, to test these findings in other settings. A final limitation is that models in the present study specified only a small number of relevant influences on safety behavior. Previous literature indicates that age, gender, overtime hours worked, and time on the job can have effects on safe work behaviors, over and above the constructs included in our models. To address this potential limitation, we conducted a post hoc test to compare a model that included these control variables as predictors of safe behavior in the system–person sequential model. Ž . The resulting model n s 320 was significantly Ž weaker than a model without these variables with control variables, x 2 s 380.4 , df s 182, CFI s 0.94; without control variables, x 2 s 303.89 , df s 2 . 128, CFI s 0.95; SCDT x s 76.54 , df s 54 . In the revised model, none of the added control vari- ables had a significant effect on safe behavior, and the direction and significance of the original path estimates remained unchanged. This test provided additional evidence for the robustness of the influ- ence of sociotechnical factors on safe behaviors. In sum, we believe that the dearth of theory development and testing in the safety arena, our focus on scale development and dimensionality, tri- angulation of results through additional sources, and our interest in formulating a new model, all support our methodology and our chosen sample. Future research can expand upon our methods by including multiple organizations, monitoring change over time, and considering a greater range of data sources.

6. Conclusions