The analysis and investigation of this study were based on data collected by a preceding study jointly conducted by the Federation of Korean Trade Unions and Yonsei University in 2016. This study aimed to demonstrate the actual prevalence of presenteeism or absenteeism among workplaces enrolled in the federation, and understand the underlying psychosocial backgrounds of presenteeism or absenteeism. The study was based on a printed questionnaire, which contained questions for basic individual information, characteristics of enterprises, experiences of presenteeism or absenteeism, and experienced symptoms.

The questionnaire sought to determine participants' demographic information such as gender, marital status, and age as well as health-related habits including smoking history and drinking habits. To investigate their working conditions, items on total weekly working hours, occupation, type of employment, shift work, tenure at the current workplace, and working position were surveyed by the self-reported questionnaire.

The Patient Health Questionnaire-2 (PHQ-2) was used to investigate the depressive symptoms of workers [23]. It is a brief survey that was designed for use in a busy clinical setting; it helps physicians easily evaluate patients’ depressive symptoms. The PHQ-2 seeks information about the frequency of depressed mood and anhedonia over the past 2 weeks, with scores ranging from 0 (“not at all”) to 3 (“nearly every day”). Thus, the PHQ-2 score can range from 0 to 6. According to the previous study, the PHQ-2 with this cutoff had a sensitivity of 87% and a specificity of 78% for major depressive disorder, and a sensitivity of 79% and a specificity of 86% for any depressive disorder [23]; therefore, it was regarded as the optimal PHQ-2 cutoff score [24]. Additionally, another study that tried to verify the reliability and validity of the Korean version of the PHQ-2, suggested the optimal cut-off point as 3, with 91.9% of sensitivity and 100% of specificity [25]. For this study, we set the cut-off point of the PHQ-2 score as 3, and defined “depression” as scoring 3-6 points and “no depression” as scoring 0–2 points.

Although job-related stress can be evaluated according to several categories, 3 categories were investigated in this study: job demands, interpersonal conflicts, and job control. The questions for job-related stress were derived from the short form of the Korean Occupational Stress Scale (KOSS), and participants were able to choose 1 of 4 options per question. According to the researchers who collected this data, the federation requested a shorter questionnaire as participants usually refused to answer questionnaires comprising several pages. Because of this, the KOSS questionnaire was shortened and only a few aspects of job-stress were selected. Hence, only job-demand, job control, and interpersonal conflicts were investigated, leading to a limitation of this study.

Questions from the official Korean version of the Work-Productivity and Activity Impairment Questionnaire-General Health (WPAI-GH), developed, translated, and provided by Reilly Associates, were used to investigate presenteeism [26]. The Work Productivity and Activity Impairment Questionnaire (WPAI) was developed in 1993, to collect data regarding productivity loss [262728]. The WPAI-GH comprises 6 questions that seek to determine the extent to which general health conditions affect personal work productivity and daily life. Six questions of WPAI-GH ask the patient the number of hours missed from work or daily activities in the past 7 days [27]. Respondents are asked to indicate the extent to which their health problems have affected their work productivity on a scale of 0 to 10 that represents impairment percentages, higher numbers reflecting more impairment and loss of productivity [27]. For this study, the existence of presenteeism was indicated when participants reported that they experienced presenteeism during the last one week and rated their working impairment a score of more than 1 point. On the other hand, the absence of presenteeism was indicated when participants reported that they had not experienced loss of productivity due to health problems in the last one week and if they rated their work impairment as 0 [2129].

In this study, all statistical analyses were performed using the SAS 9.4 program (SAS Institute, Cary, NC, USA). Demographic and workplace-related variables were compared between those 2 groups (presence of presenteeism vs. absence of presenteeism) by χ² tests or Fisher's exact tests. For this study, participants were clustered by each enterprise; therefore, the database has a hierarchical structure necessarily. Thus, we used multilevel analysis to evaluate the differences derived from the differences among enterprises. Multilevel models can be useful tools when using data with a hierarchical or clustered structure as they allow simultaneous examination of group-level and individual level factors [3031]. Usually, people exist within organizational structures, like families or schools. Similarly, workers exist within businesses, sectors of economy, as well as geographic regions. These hierarchical structured data can present some problems, because individuals who exist within one group tend to be more similar to each other than individuals who were randomly sampled. Also, people in the same group tend to share certain characteristics, in other words, observations based on these individuals are not fully independent. This characteristic of hierarchical data can lead to a higher probability of rejection of null hypothesis than if the data included truly independent observations; therefore, using multilevel analysis is needed for analyzing hierarchical data to get conservative results [32].

The intraclass correlation coefficient (ICC) was calculated by the ratio of the between-cluster variance to the total variance [33]. It indicates the correlation among observations within the same cluster. In the previous studies, when the ICC value was higher than 0.1 (10%), different results were found between standard logistic and hierarchical modeling [3435]. For this study, the approximation of ICC by Snijders and Bosker was used: $\mathrm{I}\mathrm{C}\mathrm{C}=\frac{{\mathrm{\tau }}_0}{{\mathrm{\tau }}_0+\raisebox{1ex}{${\mathrm{\pi }}^2$}\!\left/ \!\raisebox{-1ex}{$3$}\right.}$ where τ₀ is the estimated variance of the random effect of enterprises on the mean and π is the quantity 3.14159 [3637]. In our study, calculated ICC was 17.88%, and it was higher than 10%. Therefore, we used hierarchical modeling rather than standard logistic modeling.

For hierarchical generalized linear modeling, we used PROC GLIMMIX statement for the investigation [38]. The significance level was defined as p < 0.05. For multivariable analysis, general characteristics (gender, age, marital status, and smoking and drinking habits), work-related characteristics (permitted sick leave, tenure, job position, occupation, shift work, employment type, and weekly working hours), and job-related stress (job demands, interpersonal conflict, and job control) were used to make adjustments. Additionally, we stratified participants according to the presence of depression, and analyzed the relationship between job related stress and presenteeism.

Printed questionnaires were distributed to the chiefs of labor unions in each enterprise, who recruited participants from their workplaces with their written consent. In total, 30 printed questionnaires were distributed to the 60 enterprises that had joined the Federation. The datasets for analysis were built and provided by the Federation of Korean Trade Unions without any personally identifiable information. The protocol for the current study was reviewed and approved by the Institutional Review Board of Yonsei University Health System (IRB No. Y-2018-0007).