* Review the current status of research showing that reducing modifiable health risks leads to reduced absences and increased work productivity.
* Summarize the new findings on the relationship between changes in health behaviors and illness- and family-related absences from work.
* Identify the health changes most likely to reduce work absences, including their association with changes in health status.
Modifiable health risks such as smoking, obesity, poor nutrition, and physical inactivity have well-established links to preventable diseases that degrade quality of life and shorten life spans (eg, diabetes, coronary heart disease, and hypertension). This partly explains why the medical profession, the educational system, governments, workplaces, and other institutions increasingly promote public health through “wellness” and “well-being” programs and informational campaigns. External actors that pay all or most of the bills for medical treatments (eg, taxpayers, insurance companies, charitable care providers, employers, insured patients who subsidize care for the uninsured) also stand to gain from the adoption of healthier lifestyles, as do entities that benefit from individuals' labor power—primarily employers but also governments in their capacity to borrow against future economic growth. Research consistently links health risks such as body mass and smoking to productivity outcomes such as illness absences, on-the-job injuries, and reduced job performance.1–14 Several studies estimate that the productivity costs of preventable illness exceeds the medical and pharmaceutical expense of treatments.15,16 In this respect, both employers and employees share an interest in improving the overall health of the workforce.
On the other hand, the proposition that reducing health risks can lead to higher productivity is a bit more difficult to support empirically. Ideally, the relationship between changes in health risks and productivity would be assessed through randomized-controlled trials. In most instances, however, this is not practical and much of the existing literature looks at worker productivity, health, and risks cross-sectionally. As such, observed results may be spurious, due instead to workers' unmeasured characteristics. Existing longitudinal studies tend to show no relationship between health risks and reduced productivity but are either nonrepresentative,17,18 use a cross-sectional measure of absence with a differenced measure of health risk,17 or use retrospective reports of time since the change in health risk (eg, smoking).19
In addition, illness absence is frequently measured ambiguously as “sick leave”—a payroll or human resources category of absence that does not necessarily indicate illness as a cause—or without attempting to distinguish illness from other incidental absences at all. Doing so overlooks any behavioral links to absences that are not because of health but instead reflects some predisposition toward health risks and work attendance more generally.
In this article, we focus on the link between modifiable health risks—body weight, exercise, and smoking in particular—and days for which employed persons missed work because of their own illnesses. We, therefore, provide a partial test of the assertion that employers stand to realize productivity gains by promoting healthier lifestyles among their workforce (although a full test would also consider the value of improved job performance,20,21 for which measures are not available in our panel of data). Much of the existing literature on the correlates of health risks and absence rely on results from cross-sectional analyses and, therefore, are prone to bias from unmeasurable characteristics of individuals with different risk patterns or different motivations to work through illness. As a corrective, we use a first-difference approach that looks at the temporal change in health risks and illness absence among a longitudinal panel of employees. We also differentiate between absences taken for a worker's own illnesses and that taken for the illnesses of family members and estimate regression models for each separately. Our results generally corroborate the association between healthier habits and fewer days of illness absence—although without knowledge of whether employees participated in wellness programs or estimates of what it could cost employers to help workers achieve similar results relative to the wages and opportunity gains of better attendance, we sidestep the issue of whether our findings vindicate the “business case” for improved workforce health (and without information on job performance, we cannot assess the full value of productivity gains). We found no evidence that employees with health risks have a predisposition for missing work more generally. The results further corroborate the direct links between risks, health, and illness absences.
Considering the empirical links between baseline measures of employee health, modifiable health risks, and absence, we develop and test the following hypotheses.
Hypothesis 1: The change in work absences because of a respondent's own illnesses will be negatively associated with the change in health status.
H1: β > 0; H0: β = 0;
Hypothesis 2: The change in work absences because of a respondent's own illnesses will be positively associated with the change in high-risk indicators (ie, body weight and tobacco use) and negatively associated with the change in low-risk indicators (ie, frequencies of heavy and light exercise).
H2a(bodyweight and tobacco): β < 0; H0: β = 0;
H2b(exercise): β > 0; H0: β = 0;
Hypothesis 3: Because we propose that health risks influence absence by impacting health status, adjusting for changes in health status will mediate the associations between changes in illness absences and health risks.
H3: |βu| > |βs|; H0: |βu| = |βs|, where βu is the coefficient for the unsaturated model, and βs is the coefficient for the saturated model.
Hypotheses 4, 5, and 6 mirror Hypotheses 1, 2, and 3, respectively, with the substitution of change in work absence because of others' illness as the dependent variable. Generally, a rejection of the null hypothesis that β = 0 or |βu| = |βs| would suggest that the observed associations between health or health risks and illness absences reflect some unmeasured tendency to miss work that is not strictly fixed over time (eg, changes in workplace policies regarding paid-time-off or changes in workplace tempo or climate that discourage or encourage absences for legitimate reasons). Because the hypotheses specify a direction of the coefficient, we interpret statistical significance using one-tail of the T distribution.
The data come from the Panel Study of Income Dynamics (PSID).22 The PSID is a longitudinal survey that began in 1968 with a representative sample of almost 5000 US households. Within these households, demographic, economic, health, and labor force information were gathered on approximately 18,000 individuals. Since 1968, information has been collected on these individuals on an annual or (beginning in 1999) biennial basis, and new households and individuals have been added to the survey sample, because members of the original sample marry, set up new households, or have children. The original study deliberately over-sampled low-income households, which resulted in a sizable subsample of African Americans. A representative sample of more than 2000 Latino households was introduced in 1990.23 As of the 2009 wave, the PSID included information on about 22,000 individuals from more than 9000 households.
In each wave of the survey, a household member serves as the respondent for the entire family unit and the majority of questions gather information on the household head (a man in instances where an adult couple in the household is married but a man or woman otherwise) and his wife (or cohabitor), if any.
SAMPLE AND DEPENDENT VARIABLES
We limit the analysis to employed adult heads and spouses who acted as the respondent in a survey year and who had nonmissing responses to questions about whether they missed any work in the prior year because of their own illness. Our final sample in the regression models includes 6383 responses for 2917 unique individuals.
Respondents who answered the work illness absence question affirmatively were subsequently asked about how much time they missed in the prior year because of their own illness. Responses were recorded verbatim in days, weeks, or months. The follow-up questions about the frequency of missed time have been included in the PSID since 1994. Nevertheless, before 2003, the questionnaire employed a slightly different skipping logic for the question depending on whether the informant was currently employed at the time of the survey. For that reason, we limit our analysis to survey waves in 2003, 2005, 2007, and 2009. The survey asked a nearly identical question about work missed “because someone else was sick,” and pending an affirmative response, durations of absences are recorded on the same basis.
We convert all responses about missed time because of illness into workdays. Respondents answering that they missed no work because of illness were recoded to 0 and days are recorded verbatim. Responses in weeks are multiplied by 5 workdays per week; responses in months are multiplied by 21.5 workdays per month (using PSID's guidelines instructing interviewers to treat a month as equivalent to 4.3 weeks).
Recall error is an obvious concern in any retrospective study of experiences, with workers likely underreporting their sick day absences.24 Some of the gap may result from differences in how employers and employees experience work absences. From the employer's perspective, a sick day is a payroll or human resources category of legitimate reasons for which a worker misses work and may explicitly permit nonillness absences (such as doctors' appointments for one's self or a family member); from an employee's perspective, a sick day is an absence for which work was missed because of illness. Comparing survey responses of sick days and total leave time to time card records, one study found a closer correspondence in 4-week and 3-month intervals, with intraclass correlations lower for the week period (0.68 compared with 0.86) but still with good agreement.25 Studies that assess the validity of survey responses to illness absence questions over a 12-month recall generally indicate that while shorter periods are preferred, the year time frame remains valid for research purposes.24–27
Health Status and Health Risks
We use a common single-item question to measure self-rated general health as “poor,” “fair,” “good,” “very good,” or “excellent.” This question has been shown to have good reliability and validity for assessing physical and emotional health and is a good predictor of mortality and morbidity across various settings and subsamples.28,29
We use three measures of health risk—body mass, smoking, and physical activity—for which the PSID collects information and that accord with the ongoing goals of the US Department of Health and Human Services' “Healthy People” initiatives,30 given their known associations with illness and disease. Body mass index (BMI) measures the ratio between a person's weight (in kg) and his or her height (in m2). Although there are problems with using BMI as an indicator of fitness (eg, in the case of very muscular individuals), generally the incidence of disease and health complications increases with body mass. For most people, a BMI score in the range of 19 to less than 25 indicates “normal” body mass, 25 through less than 30 indicates overweight, and 30 and above indicates obesity. The obesity category is further broken down into three subcategories I (BMI, 30 to 34), II (BMI, 35 to 39), and III (BMI, 40 or more). We construct our measure of BMI using responses to questions about height and weight and transform the resulting score by its natural log.
Smoking behavior is captured by questions asking whether a respondent smoked cigarettes and, given an affirmative answer, how many cigarettes he or she smoked on an average day. Most respondents (about 80%) did not smoke in a given year, and the mode smokers (about one third) indicated that they smoked 20 cigarettes (roughly a pack a day). Given these results, we recoded the daily cigarette count ordinally to indicate that a respondent was a nonsmoker, smoked less than a pack a day, or smoked a pack a day or more.
The PSID respondents are asked how often they participate in “light physical activity—such as walking, dancing, gardening, golfing, bowling” and also in “vigorous physical activity or sports—such as heavy housework, aerobics, running, swimming, or bicycling.” For each question, responses are recorded in the number of times after a verbatim unit of time (eg, per day, week, 2 weeks, month, or year). For each type of physical activity (eg, light and heavy), we convert all responses into the number of times per week and recode the responses ordinally to indicate that a person “never exercises,” exercises “once a week,” “a few times a week” (ie, between two and six times), “once a day” (ie, seven times per week), or “more than once a day.”
Of course, illness can go only so far in explaining whether and when people stay home from work. In a very real sense, employees “choose” to work when sick or not,31,32 and their choices are influenced by social norms,33,34 job demands,35 evaluations of organizational justice,36,37 and workplace policies such as working time arrangements.38 Personal characteristics such as marital status,10 parenting,39 and socioeconomic status40–42 also play a role. Although the survey contains information on variables that frequently play a role in explanations of both health and illness absences—such as age, sex, race, occupation, and chronic health conditions such as asthma, diabetes, hypertension, and emotional problems—the estimation method used in this study largely differences the effects of these variables from the models (see later). We summarize the demographics in this study only for the purposes of describing the sample respondents. Nevertheless, time-variant characteristics that hypothetically could impact both health and illness absences are included in the models where they are available. These include family income, marital status, education level, and the presence of young children in the family. Family income is the sum of household members' taxable income, transfer income, and social security, transformed by its natural log. Marital status is dichotomized to indicate whether a respondent was married or cohabitating at the time of the interview (where 1 = yes and 0 = no). The presence of young children is dichotomized to indicate whether at the time of the survey, the age of the youngest child in the family was less than 6 years (where 1 = a child younger than 6 was present and 0 = no child younger than 6 was present, including families with no minor children). Education is an ordinal variable indicating that a respondent's maximum educational attainment was less than a high school diploma, some college, or at least a 4-year college degree. Summary statistics for the sample of individuals (as observed in the year in which they first entered the analysis) are shown in Table 1.
In principle, the relationship between changes in health risks and illness absences could be assessed, using an ordinary least squares regression equation:
Equation (Uncited)Image Tools
where y is the number of sick days for the ith person, β1 is the regression coefficient for each variable x, β0 is an intercept term when values of x = 0, and u is an error term.
Personal characteristics that are unmeasured or unmeasurable but correlated with either illness or absence could bias the results.43 Specifying the ordinary least squares model to include observable characteristics (such as chronic illness conditions) could reduce the potential for omitted variable bias, but unobserved heterogeneity among individuals with different body weights or exercise and smoking habits may remain a problem. In addition, while longitudinal panel data may produce more reliable measurement of variables than observations of a single point in time, the ordinary least squares assumption of independent observations may be violated. The error term will therefore include the influence of some variables that do not change over time and some that are time variant.43,44 We can specify this model as:
where t denotes the time period, d2t is a dummy variable equal to 1 for time period 2, a denotes time-constant error, and u denotes time-varying error. Assuming that uit is uncorrelated with xit, the model is biased and inconsistent to the extent that ai and xit are correlated. Because we are observing the same individuals over time, we can remove the bias introduced through the correlation between ai and xit by differencing Equation 2 for the each observed time period.
Equation (Uncited)Image Tools
In this “first difference” equation, Δ denotes the change in variable values from one time period to the next, and the intercept δ0 becomes the change in time periods (additional intercept terms are specified when modeling the changes over multiple time periods). Because unobserved factors in ai are constant, their influence is removed from the model.
Taking the difference in responses from any two consecutive survey years, the model specified in this study becomes
Equation (Uncited)Image Tools
Given the descriptions of the independent variables earlier, the value of the change in the ordinal and dichotomous variables is positive when an individual newly adopts the described behavior or status (eg, becomes newly married, exercises once per day or only once per week as opposed to having never exercised previously, rates health as good as opposed to rating it previously as fair); the value is negative when an individual newly discontinues the described behavior or status (eg, has no children younger than 6 years at home when previously there was at least one, stops smoking). The change in the natural log of BMI and family income are measured as interval variables measures, where negative values indicate a decrease from the previous survey and positive values indicate an increase. Regardless of the level of measurement, a variable takes on a value of zero when there is no change in status. The means, standard deviations, and confidence intervals (CIs) of the first-differenced variables are shown in Table 2.
To accommodate the multiple year panel structure of our data, we specify a generalized least squares model, with the assumption that individual effects are uncorrelated with the x variable (ie, a random effects specification). A Hausman test of differences in the model coefficients, under the assumption that individual effects were correlated with the x variables, indicated that the random effects specification produces relatively efficient results (P = 0.344). In spite of their attractive econometric properties, first-differenced models may still exhibit autocorrelation and heteroskedasticity (both across and within panels). A Woolridge test for panel autocorrelation returned a not significant result (P = 0.177), indicating that autocorrelation was not a problem in our data. A likelihood ratio test when specifying a feasible generalized least squares model with and without panel heteroskedasticity indicated that the errors were not constant. We, therefore, specify robust standard errors. All models are estimated using Stata, Software Version 12 (Stata Corp, College Station, TX).
As shown in Table 1, the average worker missed about 3.4 days per year because of his or her own illnesses, compared with about 1.5 days for someone else's illness. Most workers rate their health as at least “good,” although fewer than one in three meet the criteria for a healthy body weight (the mean BMI, 27.9, would be categorized as overweight). By contrast, three quarters are nonsmokers, nine in 10 do light exercise at least once a week (with a mode of once a day), and three in five do heavy exercise at least once per week (with a mode of “a few times per week” among workers who do any heavy exercise).
Table 2 summarizes the first differences (between 2 consecutive surveys, administered at 2-year intervals) in the variables used in the regression models. The average observation of illness absence was roughly a 15% increase from the initial baseline average; the mean change in workdays missed because of one's own illness was 0.5 days (with a range estimate from 0.1 to 0.9 days). Days missed for others' illnesses were about 4% of the baseline average, but the range included both negative and positive values. Health status and frequency of light exercise declined, while mean BMI increased. The other observed risks changed in more benign ways; the frequency of heavy exercise increased, while smoking decreased.
Table 3 shows results for random effects regression models predicting the 2-year change in workdays missed because of a respondent's own illnesses and the illnesses of others. Model 2 supports Hypothesis 1: respondents whose reported heath at t2 was better than at t1 had 1.05 fewer illness absences for each category of improvement (90% CI = −1.60 to −0.48 days; P < 0.01, one-tailed) than if their health had stayed the same. The lack of significance at the 0.05 or 0.10 levels for the coefficient for Δhealth status in model 4 corroborates the proposed relationship between health and illness absences, as opposed to absences more generally. There is no tendency for estimated absences because of others' illness to decline as one's own health improves (β = −0.13; 90% CI = −0.36 to 0.11 days; P = 0.19, one-tailed).
The results in model 1 and 2 partially support Hypothesis 2. As seen in model 1, the observed change in illness absences increases with the observed change in the natural log of BMI. As an example, on average, a 10% increase in BMI would increase estimated illness absences by about 0.44 days (ie, 4.67 × ln (1.1) ≈ 0.445; 90% CI = 0.02 to 0.9 days; P = 0.043, one-tailed). Increases in the frequency of both light and heavy exercise are associated with decreases in estimated illness absences, albeit without significance for heavy exercise. Respondents who reported doing more light exercise at t2 than at t1 had 0.33 fewer illness absences for each category of improvement (90% CI = −0.7 to 0.02 days; P = 0.06, one-tailed) than if their exercise level had stayed the same, while an improvement in heavy exercise was associated with 0.22 fewer illness absences (90% CI = −0.53 to 0.08 days; P = 0.12, one-tailed). The coefficient for Δsmoking was opposite of the hypothesized direction but also without statistical significance (P = 0.15, one-tailed).
The nonsignificant coefficients for Δlog (BMI) and Δexercise in model 3 corroborate the proposition that changes in behaviors impact absences insofar as these changes influence health but have no impact on absences for others' illnesses. Change in educational attainment was the only factor with a significant association with changes in absences for others' illnesses.
Nevertheless, evidence supporting Hypothesis 3 remains mixed. Adding the variable Δhealth status in model 2 significantly reduces the coefficient for Δlog (BMI) from 4.67 to 3.66. Nonetheless, the coefficient for Δlog (BMI) remains significant at the 0.10 level in model 2 (one-tailed). This finding suggests that much of observed association between changing body mass and illness absences reflects some unmeasured but time-varying factor. Nor does modeling the change in health status appreciably alter the association between changes in light exercise and illness absence observed in model 1.
On the other hand, including all of the Δrisks in the model simultaneously abstracts away from the reality of how people change their behaviors and, therefore, may mask the ways in which exercise, smoking cessation, and weight loss are complementary with one another. Table 4 shows the β coefficients for each risk, when entered individually into models 1 and 3 (coefficients for changes in personal characteristics were also modeled but are not shown). Only for heavy exercise does the coefficient for predicting the change in days missed for one's own illness differ appreciably from the results in model 1 (Table 3). It is reasonable to conclude that some of the negative association between heavy exercise and absence is mediated by body weight and participation in lighter types of exercise. The converse is not necessarily true, nevertheless. The coefficient for Δlog (BMI) is only slightly larger than that observed in Table 3, and the differences are not statistically significant. Smoking remains not significant, as do the relationships between health risks and absences for others' illnesses.
Taken together, the results lend support to the proposition that modifiable health risks—exercise and body weight in particular—contribute to one's capacity to attend work consistently. It, therefore, corroborates some findings linking body weight to illness absence,6,8 but which typically rely on nonrepresentative samples and cross-sectional frameworks. The representative sample and longitudinal perspective used here provide some additional context to other studies showing limited or no correlation between changes in absence, body weight, and exercise5,6,17,18 but does not lend support to the suggestion that smokers could improve their attendance by quitting.
How the relationship between risk levels and absence operates through the mechanism of health remains less clear. On one hand, changing one's risks patterns had no discernible relationship to increases or decreases in absences for others' illnesses. This suggests that health risks are not associated with a general propensity for absences but with personal illness absences specifically. On the other hand, changes in self-reported health status—itself relatively explanatory of changes in absences—only partly explained the association between BMI and illness absence and none of the association between light exercise and illness absences.
Moreover, the results for BMI and exercise in models 1 and 2 must be interpreted within the saturated context of the equation, that is, the coefficients for observed changes in BMI reflect individuals' experiences controlling for changes in the frequency of physical activity (if not the quality) and vice versa. For example, the observed impact of changes in body weight could reflect changes in nutritional habits, which could have the supplementary effect of reducing symptoms of episodic illness (such as fatigue) for which work is missed. These episodic events may be less of a factor in influencing self-assessments of health than chronic illnesses for which a person may be under ongoing medical care. A similar explanation could be posed for exercise, which could improve circulation or energy even in the absence of appreciable weight loss.
Nonetheless, several limitations to the study should be noted. Most notably, the duration between surveys may be suboptimal. Body weight and exercise behaviors may either go up or down over 2 years, which may obscure any short-term durations. On the contrary, changes in health risks may take longer than 2 years for health effects to manifest to the extent that they might impact absence. The results may have differed had the survey used an absence question with a recall period shorter than 12 months. We are unable to determine from the data whether employees participated in any wellness programs, in the workplace or otherwise. In spite of what is known about workplace climate's impact on both illness and absence, the survey does not contain questions with which climate and its changes can be measured sufficiently. Such efforts would be conducted within populations of workers from within the same workplace and preferably in the wake of “natural experiments” such as changes in leadership or mergers with other companies or after the introductions of new health promotion and wellness efforts with the explicitly purpose of improving behavior and deepening the organizational commitment for health. These studies should pay particular interest to the pathways between workplace climate and health risks because they impact both health and absences. For more targeted efforts to reduce absence through health improvements, the general population perspective taken in this study may be inadequate; focusing on specific populations (such as sedentary or overweight individuals) may produce different results.
Although the results of this study certainly do not prove the “business case” for workplace promotion of healthier behaviors—doing so would require knowledge of wages, program costs, reductions in employers' share of health care expenses, improved job performance, etc—they do support the underlying proposition that employers could experience better attendance and lower labor input costs with a healthier workforce. This occurs most clearly in the case of employees who are eligible for paid sick days—on the wage equilibrium assumption that “normal” wages are discounted in anticipation of all leave time being taken so that up to a point, every unused sick day represents a cost-free day of productivity to employers.45 But incidental absences also impose opportunity costs on many employers in the form of lost revenues, substitute workers, overtime paid to coworkers, loss of team efficiency, and so on.45,46 Not only might workplace wellness programs designed to promote and support individuals' efforts serve the long-term interest of employers, but given that employers and workers alike could benefit from the adoption of healthier lifestyles, health may be one of the rare issues in which their interests are aligned unambiguously.
1. Andrea H, Beurskens AJ, Metsemakers JF, van Amelsvoort LG, van den Brandt PA, van Schayck CP. Health problems and psychosocial work environment as predictors of long term sickness absence in employees who visited the occupational physician and/or general practitioner in relation to work: a prospective study. Occup Environ Med. 2003;60:295–300.
2. Ault RW, Ekelund RB Jr, Jackson JD, Saba RS, Saurman DS. Smoking and absenteeism. Appl Econ. 1991;23:743–754.
3. Boot CRL, Koppes LL, van den Bossche SN, Anema JR, van der Beek AJ. Relation between perceived health and sick leave in employees with a chronic illness. J Occup Rehabil. 2011;21:211–219.
4. Finkelstein EA, DiBonaventura MD, Burgess SM, Hale BC. The costs of obesity in the workplace. J Occup Environ Med. 2010;52:971–976.
5. Goetzel RZ, Carls GS, Wang S, et al. The relationship between modifiable health risk factors and medical expenditures, absenteeism, short-term disability, and presenteeism among employees at Novartis. J Occup Environ Med. 2009;51:487–499.
6. Henke RM, Carls GS, Short ME, et al. The relationship between health risks and health and productivity costs among employees at Pepsi bottling group. J Occup Environ Med. 2010;52:519–527.
7. Kuhnen AE, Burch SP, Shenolikar RA, Joy KA. Employee health and frequency of workers' compensation and disability claims. J Occup Environ Med. 2009;51:1041–1048.
8. Laaksonen M, Piha K, Martikainen P, Rahkonen O, Lahelma E. Health-related behaviours and sickness absence from work. Occup Environ Med. 2009;66:840–847.
9. Lamers LM, Meerding WJ, Severens JL, Brouwer WB. The relationship between productivity and health-related quality of life: an empirical exploration in persons with low back pain. Qual Life Res. 2005;14:805–813.
10. Leigh PJ. Correlates of absence from work due to illness. Hum Relat. 1986;39:81–100.
11. Lundborg P. Does smoking increase sick leave? Evidence using register data on Swedish workers. Tob Control. 2007;16:114–118.
12. Marmot M, Feeney A, Shipley M, North F, Syme SL. Sickness absence as a measure of health status and functioning: from the UK Whitehall II Study. J Epidemiol Community Health. 1995;49:124–130.
13. Østbye T, Dement JM, Krause KM. Obesity and workers' compensation: results from the Duke Health and Safety Surveillance System. Arch Intern Med. 2007;167:766–773.
14. Wynne-Jones G, Buck R, Varnava A, Phillips C, Main CJ. Impacts on work absence and performance: what really matters? Occup Med. 2009;59:556–562.
15. Fendrick AM, Jinnett K, Parry T. Synergies at Work: Realizing the Full Value of Health Investments. San Francisco, CA: Integrated Benefits Institute/Center for Value-Based Insurance Design; 2011.
16. Goetzel RZ, Long SR, Ozminkowski RJ, Hawkins K, Wang S, Lynch W. Health absence, disability, and presenteeism cost estimates of certain physical and mental health conditions affecting U.S. employers. J Occup Environ Med. 2004;46:398–412.
17. Serxner SA, Gold DB, Bultman KK. The impact of behavioral health risks on worker absenteeism. J Occup Environ Med. 2001;43:347–354.
18. Finkelstein EA, Linnan LA, Tate DF, Leese PJ. A longitudinal study on the relationship between weight loss, medical expenditures, and absenteeism among overweight employees in the way to health study. J Occup Environ Med. 2009;51:1367–1373.
19. Sindelar JL, Duchovny N, Falba TA, Busch SH. If smoking increases absences, does quitting reduce them. Tob Control. 2005;14:99–105.
20. Brouwer WBF, Koopmanschap MA, Rutten FFH. Productivity losses without absence: measurement validation and empirical evidence. Health Policy. 1999;48:13–27.
21. Collins JJ, Baase CM, Sharda CE, et al. The assessment of chronic health conditions on work performance, absence, and total economic impact for employers. J Occup Environ Med. 2005;47:547.
22. Panel Study of Income Dynamics, Public Use Dataset, Restricted Use Data. Ann Arbor, MI: Institute for Social Research, Survey Research Center, University of Michigan; 2011.
23. Hill M. The Panel Study of Income Dynamics: A User's Guide. Thousand Oaks, CA: Sage Press; 1991.
24. Johns G. How often were you absent? A review of the use of self-reported absence data. J Appl Psychol. 1994;79:574–591.
25. Revicki DA, Irwin D, Reblando J, Simon GE. The accuracy of self-reported disability days. Med Care. 1994;32:401–404.
26. Johns G. Absenteeism estimates by employees and managers: divergent perspectives and self-serving perceptions. J Appl Psychol. 1994;79:229.
27. Newsome S. Predicting absence without work attitudes. Paper presented at: Annual Meeting of the Canadian Psychological Association; May 27-29, 1993; Montreal, QC.
28. DeSalvo KB, Fisher WP, Tran K, Bloser N, Merrill W, Peabody J. Assessing measurement properties of two single-item general health measures. Qual Life Res. 2006;15:191–201.
29. Idler EL, Benyamini Y. Self-rated health and mortality: a review of twenty-seven community studies. J Health Soc Behav. 1997;38:21–37.
31. Dew K, Keefe V, Small K. Choosing to work when sick: workplace presenteeism. Soc Sci Med. 2005;60:2273–2282.
32. Kremer AM, Steenbeck R. Avoidable sickness absence in a Dutch working population. J Occup Rehabil. 2010;20:81–89.
33. Dew K. Pressure to work through periods of short term sickness. Br Med J. 2011;342.
34. Wynne-Jones G, Buck R, Porteous C, et al. What happens to work if you're unwell? Beliefs and attitudes of managers and employees with musculoskeletal pain in a public sector setting. J Occup Rehabil. 2011;21:31–42.
35. Zavala SK, French MT, Zarkin GA, Omachonu VK. Decision latitude and workload demand: implications for full and partial absenteeism. J Public Health Policy. 2002;23:344–361.
36. Elovainio M, Kivimäki M, Steen N, Vahtera J. Job decision latitude, organizational justice and health: multilevel covariance structure analysis. Soc Sci Med. 2004;58:1659–1669.
37. Holmgren K, Hensing G, Dellve L. The association between poor organizational climate and high work commitments, and sickness absence in a general population of women and men. J Occup Environ Med. 2010;52:1179–1185.
38. Böckerman P, Laukkanen E. What makes you work while you are sick? Evidence from a survey of workers. Eur J Public Health. 2009;20:43–46.
39. Glass JL, Estes SB. The family responsive workplace. Annu Rev Sociol. 1997;23:289–313.
40. Evans GW, Kantrowitz E. Socioeconomic status and health: the potential role of environmental risk exposure. In: Lee PR, Estes CL, eds. The Nation's Health. Boston, MA: Jones & Bartlett; 2003:93–119.
41. Johns G. Contemporary research on absence from work: correlates, causes and consequences. Int Rev Ind Organ Psychol. 1997;12:115–174.
42. Marmot M. The influence of income on health: views of an epidemiologist. In: Lee PR, Estes CL, eds. The Nation's Health. Boston, MA: Jones & Bartlett; 2003:79–92.
43. Woolridge JM. Introductory Econometrics: A Modern Approach. Stamford, CT: South-Western College Publishing; 2000.
44. Liker JK, Augustyniak S, Duncan GJ. Panel data and models of change: a comparison of first difference and conventional two-wave models. Soc Sci Res. 1985;14:80–101.
45. Pauly MV, Nicholson S, Xu J, et al. A general model of the impact of absenteeism on employers and employees. Health Econ. 2002;11:221–231.
46. Koopmanschap MA, Rutten FF, van Ineveld BM, van Roijen L. The friction cost method for measuring indirect costs of disease. J Health Econ. 1995;14:171–189.
©2013The American College of Occupational and Environmental Medicine