Abstract
The purpose of this retrospective, longitudinal study was to assess longitudinal associations between modifiable health risks and workplace absenteeism and presenteeism and to estimate lost productivity costs. Across the 4-year study period (2007–2010), 17,089 unique employees from a large US computer manufacturer with a highly technical workforce completed at least 1 health risk assessment. Generalized estimating equation models were used to estimate the mean population-level absenteeism and presenteeism for 11 modifiable health risks and adjust for 9 sociodemographic and employment-related factors. Because patient age was highly correlated with several other variables, the analysis was stratified by age (<45 vs. ≥45 years). For all ages, poor emotional health, inadequate exercise, tobacco use, and having a body mass index (BMI) greater than 35 (all P<.05) were consistently associated with both absenteeism and presenteeism. Having a BMI over 35 and poor emotional health were associated with the largest impact in absenteeism (0.46 days) and presenteeism (4.03 days), respectively. Younger and older workers had similar associations between health risks and presenteeism; however, hypertension, blood sugar, inadequate exercise, and alcohol were associated (P⋜.01) with greater absenteeism among older but not younger workers. The results suggest that productivity loss is strongly related to emotional health and obesity-related health risks (eg, BMI, exercise) but differs by age. These findings could help prioritize preventive health programs offered by employers at their worksite health centers. Given the aging of the US workforce, keeping older workers healthy and productive will be crucial to remaining competitive in the global economy. (Population Health Management 2015;18:30–38)
Introduction
A
Although some individual risk factors such as age and sex are not modifiable, evidence from the health literature has identified several individual health and lifestyle risks, such as obesity and smoking, that affect worker productivity and are amenable to intervention. 4 –16 Consequently, many employers have invested in tools and programs aimed at decreasing such risks in an attempt to improve productivity. 17 –21 Unfortunately, the impact of such interventions has been variable, so employers are demanding better evidence that their investments are leading to tangible results. 22,23,24
Challenges in measuring health risks
A major challenge to demonstrating the effectiveness of employee health programs is the dearth of information about both productivity and personal health risks in administrative data sources. 21 In response, many employers encourage their employees to complete annual health risk assessments (HRAs). 4,25 –29 In addition to self-reported health data, many HRAs also capture biometric screening test results such as blood glucose and cholesterol or link to human resource records that can provide insights into sociodemographic or occupational risks. 4,28,30,31 Beyond collecting data, HRAs can be used to encourage healthy behaviors among employees by raising awareness of their personal risk factors and directing them to appropriate educational, wellness, and disease management programs. 30,32 –35
Challenges in measuring productivity
Although a recent review identified 21 different survey instruments in the literature that assess absenteeism and presenteeism, how to best measure productivity and translate it to economic value continues to challenge researchers. 36 The appropriateness of a particular productivity measure may depend on individuals' health conditions 37 or the type of work they perform. 2,38 For example, some survey instruments that collect productivity data were predominantly developed for or tested on patients with certain conditions, such as arthritis or migraines, and may be less applicable to patients with other conditions. 2,38,39 Additionally, in workplaces that rely largely on teamwork, workers who are sick not only decrease the productivity of their role but also reduce the productivity of their coworkers. 40 Furthermore, rates of absenteeism and presenteeism are usually inversely correlated so that, given the rate of illness remains constant, as absenteeism decreases, presenteeism increases, and vice versa. 2,41 The relative proportion of absenteeism to presenteeism also is influenced by contextual factors related to organizational culture. 2
Relationship between changing risks and changing productivity
Many of the earlier studies of productivity and health risks relied upon cross-sectional designs, which are less methodologically desirable because they only represent a snapshot in time and do not lend themselves to examining directional relationships over time. 5 –7,12,15,16 However, several recent studies have used a more robust longitudinal design, which can demonstrate the directional relationship between changes in health risks and subsequent changes in employee productivity. 8 –11 Among these longitudinal studies, reducing health risks was associated with 1%–2% reduced absenteeism and 2%–9% reduced presenteeism. 8 –11 Although all 4 studies focused on changes between 2 time periods using multiple regression, they varied in their sample populations, measurement of productivity, and assessment of risks and risk reduction. For example, Burton et al 9 and Pelletier et al 10 both modeled the reduction in 1 risk, while Shi et al 8 modeled a 5% reduction in risks. Each of these 4 studies used a different HRA tool [ie, Healthier People HRA v. 4.0 (The Carter Center of Emory University and University of Michigan's Health Management Research Center), HealthMedia® Succeed™, (WellMed Inc., now WebMD Healthcare Services Group) HRA, and Healthways', Inc. Well-Being Assessment], which leads to variations in the way both productivity and health risks were measured. Only Burton et al attempted to monetize the effects of health risks on productivity, and few studies assess changes in more than 2 periods. Consequently, Shi et al concluded that more research is needed to assess the effects of productivity over multiple years. 8
Objective and hypotheses
The purpose of this study, therefore, was to assess longitudinal associations between individual modifiable health risks (MHR) and productivity using 4 years of HRA data. The primary hypothesis was to determine if fewer risks in a population or decreases in risk status over time were associated with reduced absenteeism or presenteeism.
Methods
Design
The research team conducted a retrospective longitudinal study using 4 years of administrative and HRA data from a large US employer. Repeated measures regression using generalized estimating equations (GEEs) was used to assess the longitudinal relationship between modifiable health risks and productivity, while controlling for sociodemographic and employment-related factors. This study was reviewed and approved by an independent Institutional Review Board (Quorum Review IRB, Seattle, WA).
Sample
A large manufacturer of computer components offered a free annual HRA to its employees across the United States. The study period was September 1, 2006, through October 31, 2010. Although optional, employees were offered incentives to complete HRAs. These incentives increased over the study period, and completion rates rose accordingly from 29% for the first year to 66% for the last year. Not all participants completed an HRA every year. Data from a completed HRA were included if, in the year that the survey was taken, the respondent was an active employee (ie, not retired, not a dependent, not a spouse), was aged 18–64 years, had continuous insurance coverage, and had no evidence of pregnancy (eg, obstetric or neonatal medical costs).
Measures
The research team used 3 linked data sources in this study: HRA survey data, human resources records, and employee insurance eligibility records. The HRA data consisted of annual survey responses to the Mayo Clinic Health Assessment instrument, a survey that is in widespread use by employers. 5
Response variables
There were 2 dependent variables: absenteeism days and presenteeism days. Presenteeism (ie, unproductive time at work because of illness) was defined by 8 Work Limitations Questionnaire questions included in the HRA. 42 These 8 questions were used to derive a weighted summary score that represented the percent productivity lost related to presenteeism, accounting for job demand. This percent productivity was converted into presenteeism days by multiplying by the number of working days in a year (235 days). The HRA measured absenteeism by asking employees, “In the past year, how many work days have you missed because you were ill, injured, or needed to see a doctor?” The research team monetized both absenteeism and presenteeism days by multiplying by a hypothetical daily average salary of $500.
Predictor variables
The HRA instrument contained 178 questions and measured 11 MHR factors: body mass index (BMI), blood glucose, hypertension, cholesterol, triglycerides, emotional health, exercise, nutrition, safety, tobacco, and alcohol. In the HRA database, for every employee, each of the 11 MHRs was categorized as being “at risk” (coded 1) or “not at risk” (coded 0). The criteria for defining the “at risk” category for each health risk are described in Table 1. These criteria are similar to those reported by Kowlessar et al with the exception of BMI; in the present study the research team defined as at risk individuals who had a BMI greater than 35 (ie, Obese classes II or III, according to the International Classification system for BMI). Of the predictor variables, only the variables for BMI, exercise, nutrition, and triglycerides had missing data, and triglycerides were the only MHR for which more than 10% of the data were missing. As previously mentioned, GEE was used for this analysis, and GEE treats missing data as missing completely at random.
WHO, World Health Organization
Covariates
The research team controlled for 9 sociodemographic and employment-related factors in the analysis using individual employees' characteristics derived from the human resources and employee eligibility records. Variables in the human resources data included work unit (eg, engineering), type of work (eg, technical), and exempt status (ie, overtime or non-overtime). Employee eligibility files included an employee's region of residence (ie, Pacific region, Mountain region, or other region), insurance plan type (ie, consumer driven or traditional), sex, and date of birth. The research team initially transformed age from a continuous to a categorical variable with 4 categories (aged 18 to 34, 35 to 44, 45 to 54, and 55 to 64 years), which was further collapsed because of interactions with several other covariables. Table 2 indicates final groups for all categorical variables used in the multivariable models. One additional continuous variable was included in the models to indicate the year in which the HRA was taken (coded 1–4).
N indicates unique persons; †M is the mean; ‡ SD indicates the standard deviation; § % indicates the percent of total rounded to integer.
Analysis
To assess and summarize the distribution of the response variables and predictors, the research team used frequencies, means, standard deviations, and percentages. GEE models were used to assess the longitudinal relationship between modifiable health risks and productivity, while controlling for sociodemographic and employment-related factors. Age, sex, region, work type, business unit, work shift category, exemption status, insurance plan type, hypertension status, blood sugar level, cholesterol level, emotional health status, exercise level, nutrition status, safety practices, tobacco usage, alcohol usage, triglyceride levels, and BMI were all regressed on days of absenteeism or days of presenteeism (separate models).
GEE models are an extension of generalized linear models (GLMs), and GLMs are an extension of ordinary least squares regression. 43,44,45 GLMs assume that the units of analysis (subjects or observations) are independent, but in longitudinal or cluster data this is not the case. As a result, GEE models are needed to account for these correlated units. A key feature of the actual GEE model is that it corrects for correlated data by incorporating a correlation structure in the model that describes this correlation. GEE coefficients measure the relationship between the longitudinal development of the response variable and the longitudinal development of the predictor variables. The relationship between the response variable and predictor variables at different time points are analyzed simultaneously.
The research team used the Poisson distribution for the absenteeism model and binomial distribution for the presenteeism model. Both distributions are applicable when the response variable is a count of some phenomenon (eg, the number of presenteeism or absenteeism days). The Poisson model assumes that the mean and the variance of the response variable are equal, but when the variance is larger than the mean (overdispersion), the standard errors and test statistics must be adjusted in the model or an alternative distribution must be utilized. After testing for the best model fit for the data, the negative binomial model proved to be the best adjustment for overdispersion to model presenteeism, and the Poisson model was the best fit for modeling absenteeism. The research team tested all first order interaction terms in both models. Multicollinearity (high collinearity between the predictor variables) also was assessed but was not present in the model. To assess model fit the team used quasi-likelihood statistics. A P value of less than .05 was considered statistically significant. All analyses were performed using SAS 9.2 (SAS Institute Inc., Cary, NC). 46
Results
During the study period, 17,089 unique employees completed 27,459 HRAs that met both the inclusion and exclusion criteria. Table 2 lists all univariate statistics by year. As expected given the aforementioned increase in incentives over time, more employees completed the HRA in 2010 than the previous years of 2007, 2008, or 2009. Absenteeism (mean=1.99 days; range of standard deviation [SD]=0.75–2.88) accounted for about twice as many missed productivity days as presenteeism (mean=0.96 days; SD=0.19–0.20), but both remained consistent across the study period. The most frequent health risks were inadequate nutrition, poor safety, and emotional ill health. Conversely, the least prevalent MHRs were excessive alcohol use, tobacco use, and high blood glucose. Employees who completed an HRA were predominately male, aged 45 years or older, and lived in the Pacific region. Most employees performed technical work and worked in the engineering business unit. Enrollment in the consumer-driven insurance plan increased over the study period, from 21% in 2007 to 62% in 2010, while enrollment in the traditional plan commensurately decreased over time.
Table 3 shows the parameter estimates for the 11 health risks from the multivariable regression models for absenteeism and presenteeism. Because the age variable was interacting with several other covariate factors, the research team also stratified the model by age (ie, younger than 45 years of age vs. 45 and older). All parameter estimates represent average effects that account for within-employee differences and between-employee differences, controlling for the 9 covariables (ie, year of HRA, age, region of residence, sex, work type, business unit, shift work, insurance plan, and exempt status). For example, the regression coefficient for the BMI MHR predictor in the absenteeism model is 0.22 (Table 3), and on average it represents a 0.46 day difference or $229 difference (Table 4) between being at risk and not being at risk. More precisely, this value represents the average difference in cost for an employee at BMI risk versus the cost when that same employee is not at risk. This value also represents the average difference in cost between employees at BMI risk compared to employees not at BMI risk. In a longitudinal GEE analysis it is not possible to separate the relative contributions of these component parts. Even controlling for all these factors, 7 risks—hypertension, blood sugar, emotional health, exercise, tobacco, triglycerides, and BMI—were significantly (all P<.001) associated with absenteeism in the all ages model. The remaining 4 factors—cholesterol, nutrition, safety, and alcohol—were not significantly associated with absenteeism. Similarly, the presenteeism models included 7 significant associations with health risks (ie, emotional health, exercise, nutrition, safety, tobacco, alcohol, and BMI) and 4 nonsignificant associations (ie, hypertension, blood sugar, cholesterol, and triglycerides). However, only 4 risks in the all ages model were significantly associated with both absenteeism and presenteeism: poor emotional health, lack of exercise, tobacco use, and having a BMI greater than 35.
All models control for year of HRA, age group, region of residence, sex, work type, business unit, shift work, insurance plan, and exempt status; All regression models used a log-link function and a repeated subject indicator. The absenteeism model used a Poisson distribution and adjusted for overdispersion; whereas, the presenteeism model used a negative binomial distribution and no adjustment for overdispersion was needed.
†B=parameter estimate; §CI=95% confidence interval.
BMI, body mass index; HRA, health risk assessment.
All models control for year of HRA, age, region of residence, sex, work type, business unit, shift work, insurance plan, and exempt status. A hypothetical $500 per day cost was used to estimate productivity costs.
BMI, body mass index; HRA, health risk assessment.
For younger employees (younger than 45 years of age), only emotional health, tobacco, triglycerides, and BMI remained significantly (all P<.05) associated with absenteeism. On the other hand, all risks except BMI (P=.06) remained significantly associated with presenteeism for younger workers. For older employees (aged 45 years or older), all the same 7 risks that were positively associated (all P<.001) with absenteeism in the all ages model remained so in the model for older workers. Interestingly, one additional risk, alcohol, while not significant in either the all-ages model (P=.22) or younger employee model (P=.59), was significantly (P=.01) but inversely associated (B =−0.13; 95% CI=−0.25–0.02) with absenteeism. That is, older individuals who met the criteria for alcohol risk had lower absenteeism than individuals who did not met the health risk criteria. In fact, this association was the only parameter estimate that was significantly and inversely associated in any model. Presenteeism was significantly associated with the same factors of emotional health, exercise, nutrition, safety, tobacco, and alcohol in both the younger and older employee models.
Table 4 provides the mean absenteeism and presenteeism days from the multivariable regression models presented in Table 3. Table 4 also provides estimated productivity costs (in dollars) by multiplying day estimates by a hypothetical $500 per day cost. The difference column is simply the difference between the group averages between those at risk for a health factor and those categorized as not at risk.
The risk associated with the largest difference in absenteeism days between at risk and not at risk was BMI (0.46 days; $229), whereas, the risk associated with the largest difference in presenteeism was emotional health (4.03 days; $2015). In the all ages model, alcohol use was associated with fewer absenteeism days (-0.12 days), and high cholesterol was associated with fewer presenteeism days (-0.06 days). However, the parameter estimates for these 2 factors were not significant (P=.22 and P=.89, respectively; Table 3).
In the stratified absenteeism model, the greatest difference between at risk and not at risk for younger workers was BMI and for older workers was tobacco use. For presenteeism, the risk with the greatest difference was emotional health for both younger and older workers (Table 4). Although several negative differences were noted in the stratified results, only 1 health risk—alcohol use—was significantly (P=.01) associated with fewer absenteeism days and only among older workers (-0.24 days; -$118). However, alcohol was significantly (P<.001) associated with greater presenteeism days among both younger and older workers.
The risk with the greatest impact to total productivity was poor emotional health when combining absenteeism and presenteeism days (0.34 and 4.03, respectively), accounting on average for an additional 4.37 days (ie, $2186) per person per year (Table 4). The other top health risks to total productivity were inadequate exercise (1.11 days, $556), tobacco use (1.10 days, $549), and having a BMI greater than 35 (1.02 days, $511).
Discussion
Key findings
The study hypothesis that fewer risks in a population or changes in risk over time would be associated with reduced absenteeism and presenteeism was supported. While accounting for an individual's health changes over time and controlling for various socioeconomic and workplace factors, having a BMI over 35 was the health risk with the largest impact on absenteeism and emotional health had the largest impact on presenteeism. Four health risks were consistently and significantly associated with both absenteeism and presenteeism: poor emotional health, lack of exercise, tobacco use, and having a BMI greater than 35. Younger and older workers had similar associations between health risks and presenteeism; however, hypertension, blood sugar, inadequate exercise, and alcohol were significantly associated with greater absenteeism among older but not younger workers.
Comparisons to published research
The findings of this study are broadly consistent with the conclusions reached by other authors who have recently examined associations between health risks and productivity; however, differences in methods diminished comparability. As is consistent with several other studies, 14,47 the results confirmed that presenteeism accounts for more than double the productivity losses of absenteeism. The significance and directionality of the associations found in the regression analyses were generally consistent with those of Kowlessar et al, who used the same HRA as was used in this study but did not use a longitudinal analysis. As in the present study, Kowlessar and colleagues noted poor emotional health was especially impactful on productivity, incurring an additional 0.55 days absenteeism ($176) and 3.30 days presenteeism ($1,056) when comparing high-risk to low-risk groups. Also similar to the findings of the present study, a recent longitudinal study by Shi et al found poor emotional health to be significantly associated with absenteeism (odds ratio [OR]=1.33, P<.0001) and presenteeism (OR=3.68, P<.0001). This study also noted inadequate exercise and pain as significant associations with both productivity measures. Although a 2011 study by Lenneman and colleagues 11 did not include a general variable on emotional health, depression had the greatest independent contribution (P<.001) to impairing productivity in their stepwise linear regression.
Generalizability
The study findings may not be particularly generalizable to other populations because the data came from a single employer with a highly technical workforce, in which almost three quarters of the participants were male, and most employees came from a single region of the country. Nevertheless, the prevalence of several risks (exercise, tobacco, hypertension, and blood sugar) are within the range of those reported by other longitudinal studies, such as Pelletier et al, Lenneman et al, and Shi et al. 8,11,24 For example, the prevalence of hypertension in the present study sample (15%–20%) is comparable to Lenneman et al (11%), Pelletier et al (18%), and Shi et al (22%). 8,11,24 However, BMI prevalence in the present study (7%–9%) was considerably lower than similar studies (Pelletier et al: 63%; Lenneman et al: 32%; Shi et al: 29%) because the research team chose to focus on individuals in the high range of BMI (>35), while other studies used less restrictive criteria.
Limitations
This study has a number of limitations. First, although previous studies have shown agreement between self-reported absenteeism and actual sick days, the present study was limited to self-reported data. 48 –50 Second, although the research team received the total annual claims costs for each employee, the team did not have access to event-level claims data or electronic medial record data. As a result, the research team was unable to verify self-reported MHRs with medical/pharmacy claims or electronic medical records or to control for medical conditions not addressed by any of the 11 MHRs. Without claims, the team could not properly account for comorbid conditions, a deficiency that could bias the results. Consider, for example, an individual who was identified as being at risk for nutrition who also had asthma. This employee might be more costly and use more sick days because of their asthma; however, without claims data, there was no way for the research team to control for this diagnosed disease, so these costs would have been overly attributed to the nutrition risk. Third, the lack of information on race, income, and education status meant the research team were unable to adjust for these factors in these analyses. 22,29,51,52 Finally, although the Mayo Clinic HRA instrument complies with guidelines issued by the National Institutes of Health and by the American Diabetes Association, 5 as in most studies, the HRA instruments used in this study lack formal validation. However, in this current study, the research team used lab-reported biometrics on aggregate levels to assure the HRA results were reasonable.
Implications for Clinical Practice, Future Research, and Conclusions
Despite the limitations of this study, the research team believes that the results may be useful to employers and researchers. Given the aging of the US workforce, keeping older workers healthy and productive will be crucial to remaining competitive in the global economy. This paper details the relative impact of medical and lifestyle risks on employee productivity. The findings suggest that emotional health and obesity-related risks have the strongest association with productivity loss. Understanding the relative productivity loss by health risk may help employers prioritize which preventive health programs to offer at their worksite health centers. Although such estimates to monetize productivity loss can help employers quantify the value of keeping employees healthy, consensus is needed to standardize monetization methods. 53
In particular, the findings on the top health risks could provide a focus for the preventive health programs offered by employers at their worksite health centers. An interesting and seemingly illogical finding was that being at risk for alcohol among older workers was associated with lower absenteeism in the present study. One possible explanation for this association between low costs and alcohol risk is that moderate alcohol use may be protective for cardiovascular disease. 54 Certainly, more research could elucidate interactions among health risks.
Footnotes
Author Disclosure Statement
Drs. Kirkham, Clark, Bolas, Lewis, Fisher, Ms. Jackson, and Mr. Duncan declared the following potential conflicts of interest with respect to the research, authorship, and/or publication of this article: All authors are current or former employees of one of the 2 sponsoring organizations, Walgreen Co. or Intel Corporation. Walgreen Co. funded the study development, independent institutional review board, analysis, and publication costs. Intel Corp funded all costs associated with the acquisition of data.