An Analysis of Health Factors Affecting Employees’ Absenteeism: Influences of HDL Cholesterol and Blood Sugar Levels ()
1. Introduction
The health condition of workers is a critical issue. The International Labour Organization (ILO) [1] stated that “Occupational accidents and diseases lead to devastating impacts on workers, enterprises, and entire communities and economies.” The ILO [2] estimated that “The losses in terms of compensation, lost work days, interrupted production, training, and reconversion, as well as healthcare expenditure, represent around 3.94 percent of the world’s annual GDP.” The World Health Organization (WHO) [3] mentioned that “Work-related health problems result in an economic loss of 4% - 6% of GDP in most countries… Research has demonstrated that workplace health initiatives can help reduce sick leave absenteeism and healthcare costs for companies by 27% and 26%, respectively.” The National Center for Chronic Disease Prevention and Health Promotion of the Centers for Disease Control and Prevention (CDC) [4] in the United States described “Preventable chronic conditions as major contributors to the costs of health insurance premiums and employee medical claims. These costs are at an all-time high and continue to increase in the United States.” In Japan, it is reported that the number of work-related diseases requiring work absences of four days and more was 7844 in 2017 [5] .
Evidently, health is not only essential for individuals’ well-being [6] but is also a very important social goal and benefit for individual workers, firms, and society as a whole [7] . In other words, maintaining and promoting workers’ health will not only improve their well-being but will also lead to medical expense reduction and increased productivity [8] . Health risks are a serious burden to employers owing to the loss of productivity, and various studies have been conducted on this topic [9] - [22] .
The monetary costs have been also estimated. Nagata et al. [23] estimated that the cost of absenteeism, presenteeism (reduction of productivity due to health conditions in the workplace), and medical/pharmaceutical expenses were $520, $3055, and $1165 per person per year, respectively, in the fiscal year 2014 in Japan. Ramasamy et al. [24] reported that direct costs for class I (body mass index (BMI) 30.0 to 34.9, BMI = weight (kg)/(height(m))2), class II (BMI 35.0 to 39.9 kg/m2), and class III (BMI ≥ 40.0 kg/m2) obesity were $1775, $3468, and $11,481 per-patient-per-year (PPPY), respectively, higher than those of the reference; and absenteeism/disability costs were higher than those of the reference by $617, $541, and $1707 PPPY for classes I, II, and III obesity, respectively. Gennep et al. [25] reported that the mean annual costs of absenteeism, presenteeism, and overall work productivity for inflammatory bowel disease patients were ?738, ?478, and ?597, respectively. Ademi et al. [26] estimated AUD 5.23 billion in lost GDP, with an average of AUD 101,366 lost per person with heterozygous familial hypercholesterolemia over their working lifetime using 2017 GDP data. Cawley et al. [27] estimated that annual productivity loss due to obesity was from $271 to $542 per employee in 2016. Kocakulah et al. [28] described that “Some sources, including Statistics Canada, cite that absenteeism approximates 15 - 20 percent of payroll (direct and indirect) costs.” They also discussed several possible solutions for the companies. Gianino [29] studied economic losses in terms of employee absenteeism due to influenza and reported that the average work loss due to influenza was ?27/person.
Another question is whether investments in improving workers’ health conditions are worthwhile. Baicker [30] estimated that medical and absenteeism costs fell by approximately $3.27 and $2.73 for every dollar spent, respectively. However, Verelst et al. [31] mentioned that while workplace influenza vaccination is relatively inexpensive and convenient, the return on investment is volatile for employers. From a managerial perspective, it is important to identify the risk factors that affect productivity and invest in improving these factors. Lawrance et al. [32] proposed a decision support system to improve health and well-being in the workplace and identify groups of employees at risk of sickness absence aiming to reduce or prevent absences.
Various studies have been conducted on overweight and obese individuals. Howard and Potter [33] evaluated the relationship between obesity, obesity-related chronic health conditions, and worker absenteeism using logistic regression. They found that obesity was related to higher rates of worker illness absence after controlling for demographic, socioeconomic, occupational, health-related, and behavioral variables. Keramat et al. [34] studied the relationship between absenteeism and obesity in Australian workers. They concluded that workplace absenteeism was significantly associated with being overweight and obese. Hashiguchi [35] evaluated the effect of BMI on working health risks determined by heart rate reserve (HRR) using a logistic regression model with a 40% HRR criterion. However, BMI was not a significant risk factor.
Presenteeism is a complex problem [36] and difficult to measure properly. Moreover, Marmot et al. [37] pointed out that worker absence is a good proxy for workers’ health. Therefore, the relationships between health factors and absenteeism are analyzed using health and working record data of 1136 employees in this study.
2. Data and Models
2.1. Data
The dataset contained information on the health and working records of 1136 employees at one operational site of a large corporation. The site is located in a suburb of a major city in the north-eastern region in Japan. Most employees commute by car. Health data were obtained from annual mandatory medical checkups conducted in the fiscal year 2020. Work records include information on the work schedule, actual work hours, and employees’ absences from October 2021 to December 2021. Most employees are operators supporting end customers for clients using telephones or the Internet at an indoor operational site. To calculate absence days, we excluded paid, maternity and parental, nursing care, bereavement, auspicious, and special leaves admitted by the corporation’s regulations. The number of days of absence was calculated from days of not attending work due to disability (sick or injury) and personal reasons. For most of the employees, the total number of working days during this period was 63.
Figure 1 shows the distribution of the number of absence days. 945 (83.2%) employees had no absence days. The total number of absence days for all employees was 1604 days. The number of employees with over three absence days (more than one day per month) was 94 and their absence days were 1440 days, approximately 90% of the absence days of all employees. Hence, the corporation must reduce long-term absenteeism.
2.2. Models and Covariates
Vernekar et al. [38] reported the prevalence of overweight, diabetes, hypertension, dyslipidemia, and hypercholesterolemia among IT professionals. In addition to obesity measured by BMI, health factors related to hypertension [39] - [44] , hyperglycemia or diabetes [45] - [50] , and hypercholesterolemia [26] [51] [52] [53] [54] obtained from the annual medical check-ups are considered. The following covariates were used in the analysis.
Female (dummy variable) 1 if female and 0 if male,
Age (age of an employee),
Height (height of an employee) m,
Weight (weight of employee) kg,
BMI (body mass index) weight (kg)/(height, m)2,
SBP (systolic blood pressure) mmHg,
DBP (diastolic blood pressure) mmHg,
RedCell (number of red blood cells) 10,000 per mm3,
Hemoglobin g/dL,
GOT (glutamic-oxaloacetic transaminase) units per liter (U/L),
GPT (Glutamic-pyruvic transaminase) U/L,
GGP (γ-glutamyl transferase) U/L,
Triglyceride (serum triglyceride level) mg/dL,
HDL (high-density lipoprotein cholesterol) mg/dL,
LDL (low-density lipoprotein cholesterol) mg/dL,
B_Sugar (blood sugar) mg/dL, and
HbA1c (Hemoglobin A1c) %.
The summary of these variables is given in Table 1.
As the distribution of the absence days has a heavy right tail, we consider long-term absenteeism if an employee was absent over three days and define it as L_Absence = 1 if absence days were over three days and 0 otherwise. Of the employees, 8.3% had L_Absence = 1. The following logistic regression model was used:
SD: Standard Deviation.
(1)
where Λ is the distribution function of the logistic distribution given by
, x is a vector of a subset covariate and β is a vector of unknown parameters.
2.3. Selection of Covariates
In this study, the problem is that the number of observations is not large, and the percentage of L_Absence = 1 observations is only 8.3%. This implies that the dataset does not contain much information (i.e., eigenvalues of the Fisher Information matrix are not large enough compared to variations of covariates), and all the estimates except Female are not significant at the 5% level if all covariates are included, as shown in Table 2.
Another problem is the observation of missing values for some covariates. Some covariates may not be related to the dependent variable (in this case,
Table 2. Estimation result with all covariates.
L_Absence). If we drop observations with missing values in covariates that are unrelated to the dependent variable, information will be lost. The selection of appropriate covariates is very important. Nawata [55] showed that SBP was strongly associated with the occurrence of heart disease when the two variables were directly compared. However, the importance of SBP diminished when other covariates were included, suggesting that this relationship may be spurious. It is essential to use an appropriate method to select proper covariates. The same set of covariates should be selected independently without any ambiguity.
In this study, a procedure based on likelihood ratio statistics and the Akaike information criterion (AIC), one of the most widely used criteria in model selection, is used for the selection of covariates. This procedure can be used when missing values exist in the covariates. The cross-validation method, in which the dataset is divided into two groups, estimates the model using observations in one group (training set) and selects a model based on performances in the other group (testing set); the method is widely used in areas such as machine learning. However, this method may not be appropriate because the number of observations with L_Absence = 1 is too small.
The covariates are selected by the following stepwise procedure.
1) Let y be a dependent dummy variable that takes the value of 0 or 1. Suppose that there are
(potential) covariates
, and the numbers of observations excluding missing values are
for those variables. If no missing values exist for
,
where n is the total number of observations. First, estimate models containing only constant term; that is
for
and calculate the log likelihood,
. Note that if
,
.
2) Estimate models containing one covariate,
, and calculate the log likelihood,
, and
for
. It is well known that
is the likelihood test statistic of
and asymptotically follows
under
irrelevant of
.
3) Choose
that maximizes
. Without a loss of generality, we can assume that the first variable
maximizes
. Consider models with two variables given by
and obtain the second stage log likelihood,
,
. For
such that the number of observations becomes smaller than
, re-estimate
and calculate
again using observations without missing values. Calculate
.
4) Let
be a variable that minimizes
. Then consider model with three variables
,
, and calculate the log likelihood and select the covariates that maximizes
.
5) Repeat steps k + 1 times until
for all i. This is equivalent to minimizing the AIC. The final model becomes
.
A method based on t-tests could be an alternative. However, the t-test can yield misleading results in qualitative choices and related models, particularly when the sample size is small [56] [57] .
3. Results of Estimation
Table 3 shows the results of the covariate selection using the procedure described in the previous section. The bold values indicate the largest value of
, satisfying over 1 in step j, where j covariates are included.
In the first step, HDL was selected. Female, B_Sugar, Height, and GOT were selected using the following steps. All values of
were less than 1, and the procedure was stopped at this stage. The final model, which contains five covariates, is given by:
(2)
Table 4 presents the estimation results of this model. The estimate of Female is positive and significant at the 1% level, suggesting that that the female was absent from work more frequently than the male [58] . Concerning health factors, the estimates of HDL and B_Sugar were negative and positive, respectively, and were significant at the 5% level. This means that a higher level of HDL reduces long-term absenteeism but a higher level of B_Sugar significantly increases long-term absenteeism and therefore, they are considered important health factors.
Table 3. Values of LRji and selection of covariates.
The bold values give the largest value of LRji, satisfying over 1.
Table 4. Estimation results of the selected model.
4. Discussion
The cost of absenteeism among employees is significantly high for corporate employers. Additional replacement employees and their training are necessary to maintain the corporation’s services [14] [28] [31] [32] [59] . Moreover, investments for the reduction of absenteeism must be made efficiently from a cost-benefit perspective. The results in the previous section suggest that HDL and B_Sugar are two important factors directly related to absenteeism.
Figure 2 shows the odds ratios (ORs) and 95% confidence intervals (CIs) for HDL (high-density lipoprotein cholesterol) and B_Sugar (blood sugar). As these variables are numerical, the OR for variable
is calculated by comparing
and (
+ one standard deviation). For HDL, the OR was 0.75 with a 95% CI of 0.59 - 0.95. This indicates that increasing HDL level by 15.0 mg/dL (one standard deviation) reduces long-term absenteeism by 25%. (Because the probability of L_Absence = 1 was small, the odds ratio was approximately equal to the probability ratio.) The CDC [60] has labeled HDL “good” cholesterol. According to the guidelines of the Japan Atherosclerosis Society, the normal HDL level is 40 mg/dL or over [61] . The percentage of long-term absenteeism of 1063 employees with normal HDL was 7.8%; however, for 73 employees with HDL < 40 mg/dL, the percentage was 15.1%, almost twice as large as that of the normal HDL group. The roles and mechanisms of HDL have been studied [62] [63] [64] . Ghobadi et al. [65] mentioned that nutritional intervention in the workplace seems to be more beneficial in improving HDL through a systematic review, and de Liz et al. [66] reported a positive impact of regular consumption of some types of juices on HDL levels.
For B_Sugar, the OR was 1.20 with a 95% CI of 1.01 - 1.42. Decreasing B_Sugar level by 18.9 mg/dL (one standard deviation) would reduce long-term absenteeism by 20%. Under the current Japanese criteria, an individual is diagnosed with diabetes if B_Sugar ≥ 126 mg, prediabetes if B_Sugar is 110 - 125 mg, and normal if B_Sugar < 110 mg [67] . For details on the blood sugar distribution in Japan, see Nawata [68] . The percentage of long-term absenteeism is 7.6% for 1059 normal employees; however, it becomes 18.2% for 77 prediabetic and diabetic employees. Blood sugar levels can be controlled through lifestyle improvements such as modifying eating [69] and exercise habits. CDC [70] mentioned that “The goal is to get at least 150 minutes per week of moderate-intensity physical activity.”
The correlation between HDL and B_Sugar is low and negative with a correlation coefficient of −0.144. This means that the corporation could reduce long-term absenteeism by as much as 40% by improving HDL and B_Sugar levels. This operational site has its own company cafeteria, which is used by most employees, therefore, nutrition intervention and improvement of eating habits could be relatively easy with little or no cost. It may be worthwhile to practice nutritional intervention, especially for those with low HDL or high B_Sugar levels.
Figure 2. Odds ratios and their 95% confidence intervals for HDL and B_Sugar.
5. Conclusions
Employees’ health is an important issue. It affects not only the well-being of employees but also their productivity through absenteeism and presenteeism. In this study, the health factors affecting absenteeism were analyzed because absenteeism can be directly measured without ambiguity and were good proxy of health conditions. The health and work record information of 1136 workers at an operational site of a large corporation was used in the analysis. The costs of absenteeism are significantly high and this is a serious issue from a managerial viewpoint of the corporation.
The relationship between long-term absences (more than three days in three mouth) and health factors was analyzed using logistic regression models. Because the number of observations is not very large, no health factor is significant at the 5% level when all covariates are included. Therefore, a procedure to determine the optimal selection of covariates was used. It is based on likelihood ratio statistics and the AIC.
In the selected model, the most significant health factors were high-density lipoprotein cholesterol (HDL-C) and blood sugar levels. Corporations can reduce long-term absenteeism significantly by improving these factors. These factors can be improved by modifying eating habits. The operational site has its own company cafeteria, and nutrition intervention is relatively easy. It may be worthwhile to implement nutritional interventions, especially for those with abnormal levels of HDL-C or high blood sugar.
The results of this study are based on one operational site of a corporation. The employees were mainly operators working inside the buildings. The results may be different for other types of jobs and working conditions, such as fieldwork. We may not be able to generalize the results of this study. Analyses of different types of jobs and working conditions are necessary. Evaluation of the exact cost of nutrition intervention and improvement of eating habits is also important. These topics will be investigated in future studies.
Acknowledgements
The author would like to thank the corporation for providing this dataset. The CEO of the corporation has been supporting the study. The author is grateful for the support. This study was approved by the Institutional Review Board of Hitotsubashi University (2022C003). The data presented in this study was obtained under a data use agreement with the corporation and Hitosbshsi University. The author would also like to thank an anonymous reviewer for his/her helpful comments and suggestions.