^{1}

^{1}

^{2}

^{1}

^{3}

Dr P. Légeron, currently scientific director of Stimulus, provided the data and read the final draft of the article to check for unintended errors, but did not otherwise contribute to the article. The other coauthors and the two laboratories of the CNAM (MESuRS and CEDRIC) received no support or funding for this work. This does not alter the authors' adherence to PLOS ONE policies on sharing data and materials.

Analyzed the data: MNH KA. Contributed reagents/materials/analysis tools: KA MNH PL. Wrote the paper: MNH WD GS. Conceived and designed the methodology: GS MNH.

We develop a methodological approach to identify and prioritize psychosocial factors (stressors) requiring priority action to reduce stress levels. Data analysis was carried out on a random sample of 10 000 French employees who completed, during a routine interview with the occupational physician, a 25-item questionnaire about stress levels, as well as a questionnaire about 58 stressors grouped into 5 latent variables: job control, job context, relationships at work, tasks performed and recognition. Our method combines Importance-Performance Analysis, a valuable approach for prioritizing improvements in the quality of services, with Partial Least Squares-Path modeling, a Structural Equation Modeling approach widely applied in psychosocial research. Findings on our data suggest two areas worthy of attention: one with five stressors on which decision makers should concentrate, and another with five stressors that managers should leave alone when acting to reduce stress levels. We show that IPA is robust when answers to questions are dichotomized, as opposed to the initial 6-point Likert scale. We believe that our approach will be a useful tool for experts and decision-makers in the field of stress management and prevention.

Work-related stress has become an increasingly major occupational health issue, as it has negative effects on both physical and psychological health [

Here, we propose a methodological approach to help decision makers implement a strategy to prevent work-related stress. In particular, we develop a quantitative risk assessment method to identify a set of stressors requiring priority action to reduce stress levels in the workplace. Also, an appropriate methodological tool needs to consider multiple existing stressors and the correlation structure that might exist between them, and should thus be based on multivariate data analysis.

Simple indicators such as the Cooper index [

Here, we suggest combining Importance-Performance Analysis (IPA), a valuable graphical method for prioritizing improvements in the quality of services, and Partial Least Squares-Path modeling (PLS-PM), a Structural Equation Modeling (SEM) approach, widely applied in psychosocial research.

The article is structured as follows. First, we first describe the study data. Second, we show the limitations of the Cooper index [

The present work was motivated by the analysis of a large database belonging to

Participants filled in, during the yearly compulsory routine visit with the company's occupational physician, two optional anonymous online questionnaires risk factors. More than 80% of the employees accepted to answer them. The first was a set of questions measuring psychological health (including stress levels), and the second a set of questions measuring psychosocial risk factors contributing to stress level changes.

A Psychological Stress Measure (

These 25 items define a block of items associated with a latent variable we name: “

In order to measure psychosocial risk factors at work,

The 58 items are grouped into five blocks, associated with the following latent variables:

work context

job control

relationships at work

type of performed tasks

recognition.

A method developed by Cooper & Clarke, involving a quantitative risk assessment approach to prioritize psychosocial risks in the workplace [

Here, E is the mean value of the stressor (exposure) and C (consequences) the proportion of variance of stress level explained by the given stressor (r²) obtained from a simple linear regression.

However, the relevance of this indicator is questionable, despite its ease of use. The “risk level indicator” (1) represents neither a statistical risk measure, nor an impact measure. Indeed, the r² is obtained from a simple linear regression between stress (the outcome) and a given stressor, where the impact of the other stressors is ignored. In short, this would imply that acting on a given stressor to reduce stress levels does not impact the other stressors, which is not entirely reasonable.

We illustrate this (

In the left-hand panel, we show the case of two stressors x_{1} and x_{2} with the same variability, as well as the same correlation with y, whereas the average exposure for x_{2} is higher than for x_{1} (E_{2} > E_{1}). It follows that x_{1} and x_{2} have the same risk, since increasing (or decreasing) x_{1} or x_{2} equally would have the same influence on y. However, according to Cooper’s index, x_{2} should have a higher risk than x_{1}, which is not true. In the right-hand panel, we show the case of two stressors x_{1} and x_{2} with the same average exposure level (E_{1} = E_{2}), where variability and correlation with y are different (x_{1} is more correlated with y than x_{2}). Therefore, x_{2} has a higher risk than x_{1} because y, given x_{2}, has more variability than y, given x_{1}. However, according to Cooper’s index, x_{1} should have a higher risk than x_{2}, which is not true here.

The mean value of a stressor may be considered as a performance index. However, it is not related to any effect on stress. The slope of the regression line would be a better indicator than the correlation; however, a simple regression is not pertinent here. In the next section, we show that structural equation modeling can provide an adequate solution.

To obtain insight into data on stressors and stress level, correlation analyses were conducted using structural equation modeling (SEM), which has the advantage of providing global measures of fit for latent (subjective or unobserved) variable models [

In fact, we hypothesize in the underlying conceptual model, developed with experts at

To predict the impact of the 5 job stressors blocks (exogenous constructs) on the stress block (endogenous construct), we used PLS-PM over other SEM approaches, for two main reasons. First, it allows the development of a system of weights via several indices, when the blocks of stressors are strongly correlated, which was the case with our data. Second, PLS-PM is a better choice when the Gaussian distribution assumption is not valid, which is the case for stressors (6-point Likert scale). PLS-PM uses an iterative algorithm; after convergence, latent variable scores are obtained for each observation, and structural coefficients are estimated using multiple regression.

This approach appears to be a useful tool for psychosocial risk management policy for the workplace, and comes with two advantages. First, it allows us to build a relevant scale for stress level using the 25 items in the PSM25 questionnaire, rather than using an equally-weighted scale. Second, it gives a ranking of the five blocks of stressors according to their predictive impact on stress level, using path coefficient values.

However, in order to directly prioritize stressors impacting stress level, we considered constructing a system of levers to identify stressors which decision makers should maintain or act on in priority, to reduce stress level. To achieve this, we suggest using importance-performance analysis (IPA) [

IPA is an easily understandable graphical tool presented as a grid divided into four quadrants. The horizontal axis shows the item’s performance, and the vertical axis its importance. The four quadrants are as follows. A is the top-left quadrant (“Concentrate Management Here”), B the top-right (“Keep up the Good Work”), C bottom-left (“Low Priority”) and D bottom-right (“Possible Overkill”).

Of most interest are items in quadrants A and B, as these are relatively “more important” than those in quadrants C and D. Therefore, an item with a lower performance and a higher importance falls into quadrant A, indicating that decision makers should devote further resources to this particular attribute, so as to improve its performance.

An item's performance is defined as its observed mean score over the 10 000 individual responses to this item. Its importance is calculated with the formula (2) in which the importance of a given item is the product of the absolute value of the outer weight and the path coefficient of the latent variable in which the item belongs, obtained from the PLS-PM:

The study was approved by the French Data Protection Authority “La Commission Nationale de l’Informatique et des Libertés” (CNIL, #1839949 v 0).

To evaluate the homogeneity (or unidimensionality) of the six blocks, eigenvalues of the correlation matrix between manifest variables (stressors) belonging to the same block were calculated [

Latent variables | # items | 1^{st} eigenvalue |
Cronbach's α |
---|---|---|---|

Stress (PSM25) | 25 | 10.92 | 0.94 |

Work context | 14 | 5.55 | 0.88 |

Job control | 14 | 3.99 | 0.80 |

Relationships | 12 | 4.97 | 0.87 |

Tasks | 12 | 2.86 | 0.62 |

Recognition | 6 | 3.15 | 0.81 |

First, correlations between the five blocks of stressors (

Context | Control | Recognition | Relationship | Tasks | Stress | |
---|---|---|---|---|---|---|

Context | 1.00 | |||||

Control | 0.78 | 1.00 | ||||

Recognition | 0.72 | 0.63 | 1.00 | |||

Relationship | 0.69 | 0.67 | 0.60 | 1.00 | ||

Tasks | 0.61 | 0.72 | 0.54 | 0.53 | 1.00 | |

Stress | -0.52 | -0.63 | -0.43 | -0.51 | -0.52 | 1.00 |

Second, the measurement model allowed us to confirm the validity of the six latent constructs. As for the total stress and total stressors, the outer weights were statistically significant, indicating the relevance of the latent variables. Thus, the measurement model quality is satisfactory. Furthermore, normalized outer weights of items belonging to the stress block could potentially be used to define a more relevant stress scale, using the PSM25 questionnaire. This is a more relevant scale than that with all 25 items equally weighted as described above in the Stress measurement paragraph.

Third, the structural model allowed us to evaluate the strength and significance of the path coefficients for the relationships (structural paths) hypothesized between the constructs. The path coefficient values (

Finally, the assessment of the model’s quality shows the good ability of the model to predict the endogenous construct “stress” (R² = 0.40), indicating that the five blocks of stressors explain 40% of the variance of the stress construct (

We performed PLS-PM as described above, using a 10 000 employees dataset. A randomly drawn subsample of 5 000 employees was used to develop the models, and the remaining 5 000 employees' data served for the validation step. Analyses were performed using the PLS-PM package from the XLSTAT software (

The IPA plot of importance according to performance values for each of 58 stressors is shown in

In terms of the most important items, for the most part the five following items were identified in quadrant A, where improvement in performance is the most pressing, and upon which management should concentrate its efforts:

Task_11: « I have to work fast in a short timeframe »

Recon_01: « My promotion prospects are weak »

Recon_02: « My company offers me interesting career opportunities »

Task_09: « I work in a noisy and hectic atmosphere »

Recon_04

Of the most important items, for the most part the following six items were identified in quadrant B, where efforts should be maintained:

Task_05: « I frequently see the work pile up without being able to eliminate the backlog »

Task_07: « My work gives me many opportunities to perform interesting tasks »

Task_01: « My work has meaning to me »

Task_02: « My job involves monotonous and repetitive tasks »

Contro_03: « I can achieve professional life—personal life balance »

Contro_14: « I'm undergoing or I expect to undergo an undesirable change that might affect my career »

These results from the analyzed data suggest that, for organizational strategies to prevent and manage stress, decision-makers should act in priority on the level of the 5 “stressors to improve", and should maintain the level of the 6 “stressors to maintain”, because otherwise the stress level will increase.

IPA’s robustness against changes in the scale used to categorize the answers to the 58 items related to stressors, was tested as follows. Instead of using the 6-point Likert scale for individuals’ answers to each of the 58 items, we

Negative responses: 0 (strongly disagree), 1 (disagree) and 2 (slightly disagree) were coded 0

Positive responses: 3 (slightly agree), 4 (agree) and 5 (strongly agree) were coded 1

Results (not shown here) demonstrate a similar distribution of the items in the four quadrants. In particular, the items to improve (from quadrant A) and the items to maintain (from quadrant B) were similar. This indicates the robustness of the IPA results.

For the last twenty years, work-related stress has been recognized as a major factor in employee health and company performance. It is now widely recognized that besides biological, chemical and physical agents, the psychosocial working conditions are important determinants of employee health. Many studies have documented the effect of adverse psychosocial work factors on the incidence and prevalence of health problems [

This work aimed to develop a new statistical approach to identify a set of stressors, among many known possibilities, in order to prioritize preventive actions, using two complementary powerful statistical methods: PLS-PM and IPA. To our knowledge, this is the first time these two approaches have been used together to answer a single question. The use of this strategy provides additional insights to understanding the relationship between different stressors and stress level. Our results show that this approach could provide a more relevant diagnosis of stress predictors than Cooper’s index, or any other tool based on univariate statistical analysis.

Although several models are available to identify risk factors impacting stress level, none are based on a solid theoretical base. These models were developed using empirical field studies. However, it is very important to be able to take into account interactions between the various risk factors for stress, identified through epidemiological studies. The PLS-PM approach allows us to predict stress levels using five strongly correlated blocks built from 58 stressors, and to understand concepts that are difficult to formalize. Using path coefficients, this approach allows us to prioritize the five stressor's latent variable constructs based on their ability to predict stress's latent variable construct.

The IPA approach allows the direct identification of stressors requiring priority attention for managerial action to reduce workplace stress levels, using an indirect computation of item importance coming from the PLS-PM results. Based on a plot of item importance in relation to measured performance, IPA provides a useful and easily understandable graphical guide as to how the quadrants differ from one another. As a result, it allows decision-makers to identify areas in which they must reallocate resources [

The use of data from

As several studies suggest that women suffer more stress than men [

As correlation does not imply causality, a causal analysis should also be performed to determine the stressors on which to act in order to reduce psychosocial disorders associated with high stress level. Bühlmann proposes an analysis based on causal graphs [

We have proposed using a multivariate statistical approach based on IPA combined with PLS-PM. The results from applying this approach to data from

Our approach could be a useful tool in evaluating the impact of organizational and environmental factors on individual’s stress levels. However, it can used to study any other psychological health outcome or concept (performance, fatigue, anxiety, etc.).

(DOCX)

(DOCX)

(DOCX)

(DOCX)

We thank the reviewers for their careful reading of our manuscript and their many insightful comments and suggestions. We thank Fabien Valet, biostatistician at

The results of this study were presented in part at the 18th International Conference of Partial Least Squares and Related Methods (PLS2014), Paris, France, and at the 16th Conference of the Applied Stochastic Models and Data International Society (ASMDA 2015), Piraeus, Greece.