Analysis of the Importance for Perceived Stress using LightGBM and SHAP Values

Abstract

It is reported that 58% of workers feel severe stress while working. Mechanisms of stress are related to various factors and it was difficult to study them with conventional statistical methods because of multicollinearity. We calculated importance of 31 factors to perceived stress multilaterally with data of a big company’s data-health project. 31 factors included job stressors (branch, commute hours, job description, working hours, midnight work, work life balance, psychological safety, job demand, job control), individual factors (age, occupation, seniority, living with someone, sense of coherence (SOC; to measure stress coping ability)), non-work factors (BMI, smoking status, drinking status, exercise status, housework hours, nursing status, Pittsburgh Sleep Quality Index (PSQI; quality of sleep)), and buffer factors (supervisor support, co-worker support). A self-administered questionnaire was conducted on 7,255 male full-time employees who worked for a certain logistics company. In total, 6,166 (85.0%) of the employees agreed to participate. After we confirmed multivariate configuration factor by principle component analysis, we analyzed the data use a machine learning method, ensemble learning, which is based on a decision tree algorithm. We compared the performance of different algorithms, such as RandomForest (RF), eXtreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), and Category Boosting (CatBoost). After adjusting the learning model to calculate the importance of the explanatory variable to the objective variable, we adopted the SHapley Additive explanation (SHAP) values, which has consistently been used to interpret model findings. The LightGBM method had the highest performance, so we adopted LightGBM. Regarding importance, SOC was the most important and PSQI was second compared with the remaining variables. These findings indicate the importance of measures for improving stress coping ability and improving quality of sleep compared to job stressors like job demand. This study can be used to help prioritize steps for stress reduction in the workplace.