Doing Well by Making Well: The Impact of Corporate Wellness Programs on Employee Productivity

Abstract

This paper investigates the impact of a corporate wellness program on worker productivity using a panel of objective health and productivity data from 111 workers in five laundry plants. Although almost 90% of companies use wellness programs, existing research has focused on cost savings from insurance and absenteeism. We find productivity improvements based both on program participation and post-program health changes. Sick and healthy individuals who improved their health increased productivity by about 10%, with surveys indicating sources in improved diet and exercise. Although the small worker sample limits both estimate precision and our ability to isolate mechanisms behind this increase, we argue that our results are consistent with improved worker motivation and capability. The study suggests that firms can increase operational productivity through socially responsible health policies that improve both workers’ wellness and economic value, and provides a template for future large-scale studies of health and productivity.

Full text

1"
"
Doing Well by Making Well: The Impact of Corporate
Wellness Programs on Employee Productivity*
Timothy Gubler
School of Business Administration, University of California, Riverside
tgubler@ucr.edu
Ian Larkin
Anderson School of Management, UCLA
ian.larkin@anderson.ucla.edu
Lamar Pierce
Olin School of Business, Washington University in St. Louis
pierce@wustl.edu
Forthcoming in Management Science
June 28, 2017
Abstract: This paper investigates the impact of a corporate wellness program on worker
productivity using a panel of objective health and productivity data from 111 workers in five
laundry plants. Although almost 90% of companies use wellness programs, existing research
has focused on cost savings from insurance and absenteeism. We find productivity
improvements based both on program participation and post-program health changes. Sick
and healthy individuals who improved their health increased productivity by about 10%,
with surveys indicating sources in improved diet and exercise. Although the small worker
sample limits both estimate precision and our ability to isolate mechanisms behind this
increase, we argue that our results are consistent with improved worker motivation and
capability. The study suggests that firms can increase operational productivity through
socially responsible health policies that improve both workers’ wellness and economic value,
and provides a template for future large-scale studies of health and productivity.
Keywords: Worker Productivity, Health, Wellness Program, Presenteeism, Corporate
Social Responsibility
*We thank Hengchen Dai, Andrew Knight, numerous conference participants, and the editorial and review team for
their thoughtful comments and recommendations. Doctors Aaron Miller, Kurt Munzer, and George Toth provided
valuable medical advising. Executives and employees at LaundryCo and the program vendor generously provided time
and data for the paper. This research was supported by the Center for Research in Economics & Strategy (CRES) at the
Olin Business School. All mistakes are ours alone. Please send all correspondence to ian.larkin@anderson.ucla.edu.

2"
"
1. Introduction
Companies increasingly invest in employee health and well-being (Ton 2014). A recent survey found that
around 90% of companies use corporate wellness programs that can include simple biometric screenings such
as basic blood tests, advanced screening for diseases such as cancer, exercise programs, nutritional and diet
programs, health history and habits surveys, and training on protecting and improving health (Medical Billing
and Coding 2012). The prevalence of such programs is unsurprising given growing rates of obesity, diabetes,
and other health problems, and the implications of these issues for employer-sponsored health insurance and
absenteeism (Baicker et al. 2010; Boles et al. 2004). Obesity has steadily increased to almost 35% of the
United States population in 2012 (NCHS 2012), diabetes cases have more than quadrupled since 1980 (CDC
2014), and exercise rates and eating habits have not improved (Gallup-Healthways 2012). This decrease in
employees’ physical health is reflected in the 131% increase in health insurance premiums from 1999-2009,
the cost of which is largely borne by employers (Kaiser/HRET 2012).
Extensive research in the fields of medicine, public health, and health economics shows that the costs of
corporate wellness programs are dwarfed by reductions in insurance costs and absenteeism. A recent meta-
analysis found that each dollar spent on wellness programs saves $3.27 in health care costs and $2.73 in
absenteeism costs (Baicker et al. 2010). Although these gains are substantial, they ignore an important class of
operational benefits from investing in employee health and well-being—worker productivity. As management
and operations scholars argue (Danna and Griffin 1999; Goldstein 2003; Grant et al. 2007; Neumann and Dul
2010; Ødegaard and Roos 2014), healthier employees are not only less expensive and less absent, but are also
more productive. A large occupational health literature on “presenteeism” – working while ill – suggests lost
productivity costs firms over $150 billion, almost three times absenteeism costs (Stewart et al. 2003). Despite
this value, existing research has not causally linked objective health data from wellness programs with
measurable worker productivity changes. This may explain firm confusion about how much wellness
programs actually help their bottom line. A recent article in the Wall Street Journal notes that “employers are
stymied by the difficulties of measuring the financial and health impact of wellness programs” (Weber 2014).
In this paper, we provide the first causal evidence based on objective data that wellness programs and
their related health improvements can increase worker productivity. We use novel longitudinal data on
individual health and productivity from an industrial laundry company (which we call LaundryCo). LaundryCo
provided free annual biometric screenings to all full-time employees in four laundry plants; a fifth plant did
not participate because it offered a different insurance plan. The voluntary program increased awareness of
and attention to health by providing each employee with a personalized health packet detailing current health
status and providing recommendations for improvement. Detailed daily production data, combined with
annual medical data including bloodwork, lifestyle choices, and vital statistics, allow us to exploit this quasi-
experiment to generate causal estimates of the wellness program’s productivity impact. The program appears
to have improved average worker productivity by over 4%, approximately equal to adding an additional day
of productive work per month for each employee, although this estimate is imprecise. We find more precise

3"
"
evidence of productivity gains among those who improve their health that suggest increased work capability.
Sick employees whose health improved showed a 10% productivity increase. Strikingly, already healthy
employees who improved their health showed an 11% productivity gain. Survey and blood results indicate
that these health improvements stemmed from physical activity increases, attention to diet, and other lifestyle
changes. In addition, employees with no health problems and no health improvement showed a 6%
productivity increase following the program’s introduction. We conservatively estimate a return on
investment of 76% purely based on productivity gains, with a much higher return if the company were able to
reduce non-participation in the program and reduce turnover.
Our paper contributes to several key literatures in operations, management, and strategy. First, we add to
growing research in management on the relationship between employee well-being and organizational
performance, providing the first causal evidence linking a multi-year panel of medical data to actual individual
productivity improvements in a firm. This supports existing cross-sectional and self-reported data in prior
work (e.g., Burton et al. 1999; Goetzel et al. 2003; Ødengaard and Roos 2014).
Second, our paper joins the growing literature in operations that uses detailed micro-level production and
service data to study environmental, social, and psychological factors that impact worker productivity. Recent
studies demonstrate that operational policies such as scheduling and staffing (Chan et al. 2014a; 2014b; Dai et
al. 2015; Huckman and Pisano 2006; Huckman and Staats 2011), monitoring and transparency (Bernstein
2012; Buell et al. 2017; Pierce et al. 2014; Staats et al. 2016; Tan and Netessine 2015), performance
recognition (Gubler et al. 2016; Larkin 2011; Song et al. 2017), and workflow (KC 2013; Kuntz et al. 2014;
Staats and Gino 2012; Tan and Netessine 2014) can all affect worker performance. Even environmental
factors outside managerial control such as weather (Lee et al. 2014) may determine productivity. Our study is
unique in this literature not only by linking health and productivity, but also by supporting the argument that
firm policy can improve operations through worker health and well-being (Ton 2014).
Third, our paper contributes to the large and often conflicting literature on corporate social responsibility
(CSR) and firm performance (Margolis and Walsh 2003; Orlitzky et al. 2003; Kitzmueller and Shimshack
2012) by detailing mechanisms through which firms can “do well by doing good.” The empirical challenges in
this research are widely recognized (Barnett and Salomon 2006; Chatterji et al. 2009), prompting calls to
investigate the micro-foundations of CSR (Aguilera et al. 2007). Our study responds to this call, showing how
one frequently-used CSR policy can improve worker productivity and firm outcomes.
Finally, our paper contributes to a vast, but largely correlational, medical and public health literature by
linking objective health improvement data with objective worker productivity gains in a quasi-experimental
setting. This distinction is important for two reasons. First, people are often dishonest or biased about their
own productivity (or that of others) (Podsakoff et al. 2003). Second, self-reported productivity is difficult to
quantify or value, particularly when measured with scales.
Although our combination of objective health and productivity data provides unique evidence linking
employee health initiatives to productivity, we caution that our sample of 111 workers limits our statistical

4"
"
power. Many of our parameter estimates are only weakly significant at traditional levels, and we emphasize
that this paper should be viewed as preliminary evidence of how wellness programs improve both health and
productivity. We hope that this evidence will motivate companies with larger workforces to collaborate in
research that might more precisely estimate the effects observed here.
2. The Potential Impact of Wellness Programs on Productivity
2.1 Evidence Linking Socially Responsible Employee Policies to Performance
A large literature suggests that socially responsible (CSR) employee policies, such as wellness programs, might
improve firm performance, focusing on predictors of firm-level CSR, the outcomes of CSR, and mediators
and moderators of the CSR-outcomes relationship (Aquinis and Glavas 2012; Wang and Qian 2011). These
studies correlate CSR with outcomes that include improved reputation, consumer loyalty, and attractiveness
to investors. Despite this research, however, scholars continue to disagree about the importance of such
effects because few papers provide causal tests of CSR and firm performance (Flammer 2015).
One stream studies how employee-focused CSR policies might increase performance, finding correlations
with employee behaviors and attitudes that include organizational identification (Carmeli et al. 2007),
employee engagement and citizenship (Glavas and Piderit 2009; Jones 2010), employee relations (Agle et al.
1999), and attractiveness to job seekers (Turban and Greening 1997). Field studies almost exclusively lack
strong causal evidence, with only a few exceptions—none of which study worker health. Burbano’s (2017a;
2017b) online labor market experiments show that socially responsible messages increase effort. Carnahan et
al. (2017) uses the September 11 terrorist attack to link pro bono work to law firm retention. Flammer and
Luo (2017) links retention and engagement to CSR using shocks to unemployment insurance benefits.
This paper focuses on employee wellness programs—one of the most common socially-responsible
policies targeting employees. One reason these programs are so common is that the immediate, salient, and
measurable benefit from reduced insurance premiums provides managers with easy financial justification for
program introduction in the face of internal political opposition (Berry et al. 2010). Additionally, firms may
see immediate tangible gains from healthier employees through reduced absenteeism, injury, and worker
compensation claims (Chapman 2012). Widespread evidence shows that decreasing a health risk such as
smoking can significantly reduce insurance and absenteeism costs (Burton et al. 1999; Goetzel et al. 2003).
Existing research therefore has primarily studied firm wellness program benefits via reduced costs from
insurance, absenteeism, and risk, rather than from worker productivity. From an empirical perspective, this
focus is understandable. The link between wellness programs and productivity is difficult to causally measure.
Matched objective productivity and health data are difficult to obtain from firms, and isolating the treatment
effect of such programs amidst other policy changes can be daunting. Furthermore, most companies offer
wellness programs to all employees, which means researchers cannot untangle temporal productivity changes
caused by time trends or shocks affecting all workers. These factors may explain why meta-analyses of
financial returns from wellness programs include almost no productivity-based returns (Chapman 2012).

5"
"
2.2 Mechanisms
We argue that a corporate wellness program could affect employee productivity through two classes of
mechanisms: job motivation and capability. In the next sections, we explain that the likely importance of
motivation and capability in productivity gains depends on two observable employee characteristics—pre-
existing health problems and health improvement during the program. Although this study cannot identify
the sources of employee health problems or improvement, these individual characteristics can help illuminate
whether motivation or capability improvements likely drive our observed productivity gains.
Figure 1 shows four categories of employees likely to improve productivity through these mechanisms.
Participating employees can be categorized based on two observable characteristics across time: if the
program identified a health problem and whether the employee improved their health between exams.
Because even healthy employees can improve their health, this classification generates four participant types.
We argue that all four types might increase productivity because of motivation based on job satisfaction.
Those with health problems might enjoy further motivational gains due to reciprocity based on feelings of
gratitude. Additionally, workers who improve their health might increase productivity because of improved
physical and mental capabilities. Previously unhealthy employees who improve their health might see the
largest capability gains because of existing impediments of stress, pain, and physical weakness.
<<< INSERT FIGURE 1 HERE >>>
2.2.1 Productivity Improvement Through Motivation: Increased Job Satisfaction. One way that a
wellness program might improve motivation and therefore productivity is by demonstrating an employer’s
concern for employee well-being, which spans many psychological, physical, and social dimensions (Grant et
al. 2007). The costly implementation of a program designed to improve well-being might improve worker
attitudes by credibly signaling to employees the firm’s broader concern for the quality of their work life, and
even the quality of their life outside the workplace. The program in our empirical setting, for example, costs
approximately $240 per year for each employee.
Employers frequently offer corporate wellness programs to all employees without knowing who will
benefit. Even workers themselves may not be able to predict, ex ante, whether they will benefit. Therefore, all
participants, not just those who receive new information from the program, might perceive increased
organizational support and experience increased motivation through job satisfaction. Indeed, wellness
programs, independent of efficacy, have been shown to increase job satisfaction (Zoller 2004) and raise
perceptions of organizational commitment towards employees (Parks and Steelman 2008). Both job
satisfaction and perceived organizational support have been positively associated with job performance
(Armeli et al. 1998; Yee et al. 2008).
2.2.2 Productivity Improvement Through Motivation: Gratitude and Reciprocity. While all
employees might improve motivation via job satisfaction, employees who learn they are ill might improve
motivation more than their healthy colleagues. One of the major aims of a wellness program is to help
employees identify and focus attention on existing conditions and illnesses. Almost half of overweight

6"
"
Americans don’t recognize their weight problem (Ingraham 2016), while one in four Americans with diabetes
are unaware of their condition (CDC 2014). All such employees, regardless of whether the program also helps
them remediate their illness, might feel gratitude to the employer for providing information about existing but
unknown health conditions. Since this information is inherently valuable, employees who receive this gift will
be inclined to reciprocate (Bartlett and DeSteno 2006; Tsang 2006). Reciprocity theory holds that actors such
as employees react to unexpected giving by responding in turn, even if a receiver does not want or will not
use a valuable gift (Grant and Gino 2010; Adams et al. 2012). When an employee receives a work benefit, she
may strive to relieve this imbalance though contributions to the organization (Eisenberger et al. 2001). One
natural way employees could reciprocate is via increased productivity. Although all participants might feel
some gratitude, the value of the information provided by the program is highest for workers with preexisting
health problems, making these workers most likely to feel gratitude and thus reciprocate (Dabos and
Rousseau 2004; Hekman et al. 2009).
2.2.3 Productivity Improvement Through Work Capability. A wellness program may also increase
productivity by helping employees improve their health, thereby strengthening their work capability. As noted
above, employees might learn of an existing health condition, realizing the need to improve their health. Yet
even if they were aware of a health problem, the program may focus their attention on health in ways that
nudge them toward action. Free counseling on nutrition, substance abuse, weight, exercise, and other health
habits provided through the program may help all employees to make positive lifestyle changes through
several behavioral mechanisms. First, counseling from the program may produce the type of concrete
improvement plan known to improve health behavior (Milkman et al. 2011; Dai et al. 2012; Rogers et al.
2015; Beshears et al. 2016). Second, the program may provide an effective reminder that establishes better
habits (Calzolari and Nardotto 2017). Recent research shows that subtle nudges about lifestyle choices and
health and longevity can indeed increase commitment to healthy choices in diet, exercise, and sleep (Vallgårda
2012). The specific actions that employees take in a wellness program – seeing a registered nurse, getting
blood drawn, and receiving information on typical health problems – may provide one such nudge.
Employees who improve health problems would likely see the largest productivity gains because of
substantially improved physical capability from remediating health issues. There is consensus in the
occupational health literature that poor health reduces work capacity and substantively changes wages, hours
worked, labor force participation, job choice, turnover, retirement, and occupational choice (Currie and
Madrian 1999). There is scant evidence, however, of the direct link between health improvement and on-the-
job productivity at the individual level. Instead, researchers have indirectly connected health to productivity
either through human capital development identified by educational attainment (Becker 2007; Conti et al.
2010) or through national and other macro-level measures (Currie and Madrian 1999). But even healthy
employees without diagnosed problems might improve health through adopting healthy lifestyle choices in
areas such as diet, exercise, and sleep. Each of these has been linked to employee on-the-job productivity
through capabilities based in stamina, energy, and mood (Thayer et al. 1994). A substantial literature on

7"
"
presenteeism—present workers impeded by illness, depression, injury, or pain—links self-reported
productivity loss to mental and physical health conditions that include diabetes, depression, anxiety, cancer,
migraines, and arthritis (Stewart et al. 2003). For example, surveys from Lockheed Martin correlated
productivity losses with health problems such as migraines (4.9% loss), allergies (4.1% loss), asthma (5.2%
loss), influenza (4.7% loss), and depression (7.6% loss) (Hemp 2004). Even when health improvements are
not specifically tied to the capability to carry out a critical job task, they may increase worker task productivity
due to improved mental health and reduced distraction from pain and discomfort. Recent evidence from
Christian et al. (2015) suggests that pain reduction, and subsequent energy improvements, strongly affects
discretionary tasks such as productivity or prosociality.
Thus, while we expect that general health improvements through lifestyle choices could increase
productivity through better job capability for all workers, we would expect these gains to be largest for those
with existing health problems.
3. Data and Methodology
The major innovation in our data is the ability to link objective employee health data with daily individual
productivity measures in a quasi-experimental setting. Previous wellness program studies have either lacked a
control group or used purely cross-sectional or survey data. Our paper features both a quasi-control group
and longitudinal objective data for both productivity and health paired with self-reported survey and
demographic data for 111 workers in four treatment plants and one quasi-control plant. The major limitation
of our data, as we will detail in Section 4.2.1, is our relatively small sample, which limits our statistical power
to precisely estimate parameters.
3.1 Setting
Our data originate from a private industrial laundry service company that we call LaundryCo. LaundryCo is one
of the largest independent industrial launderers in the United States, providing uniforms, mats, and other
garments to clients that include auto shops, construction companies, restaurants, and hospitals. Items are
regularly laundered and repaired by LaundryCo using an innovative IT and production system ensuring timely
delivery. The cleaning and repair of products is carried out primarily by workers in five plants across four
states. These plants are nearly identical in layout, equipment, staffing positions, and products, and have about
15 production workers employed per laundry line on a regular day. Three of the plants, including our control
plant, have two lines with about 30 workers per day, and the other two plants have a single line with about 15
workers per day. Plants also employ other workers for whom we cannot measure productivity, including
managers, delivery workers, and support staff. The average production worker stays approximately four and a
half years, with a median tenure of two years. The primary difference between our plants is the unionized
workforce at the control plant, which reduces turnover. This difference is fixed throughout the period of our
study, and therefore will not bias our identification of the wellness program’s treatment effect.
Soiled clothes, mats, and linens are dropped off at the plants and proceed through a complex sequence of

8"
"
sorting, cleaning, drying, pressing, and repair before being loaded and returned to customers. Non-uniform
items such as floor mats or linens go through their own unique process. Workers are cross-trained on
multiple tasks, but typically specialize in only a few. Similar to other service operations, worker efficiency is
crucial to LaundryCo profitability. Mistakes and bottlenecks create costly production line disruption and leave
downstream workers with insufficient workload. Mistakes may include incorrect initial sorting, insufficient
cleaning, or failure to repair garments before the final quality check. Bottlenecks result both from mistakes
and the inefficiency of upstream workers. For instance, if the dryer operator falls behind, the pressing
machine operator might be idle until garments arrive for pressing. Downstream worker dependence on
productivity in upstream colleagues magnifies presenteeism costs for LaundryCo.
In Spring 2010, LaundryCo hired an outside company (“the vendor”) to provide a free wellness program to
employees. Management’s goal was to reduce insurance premiums and improve employee health. The
program was offered to all employees as a voluntary free benefit, with management actively encouraging
participation. Excluding managers and non-factory workers such as salespeople, 136 employees chose to
participate, while 31 chose not to. LaundryCo’s human resource director told us that employees chose not to
participate because of either a spouse’s insurance coverage, absence on the day of the program, a fear of
doctors, or worry that the program would uncover drug use (although the program did not test for drug use).
Participating employees received a 15% monthly insurance premium decrease (about $1.75 to $11 per
month), which represented about 0.4% of monthly wages for the average worker.
We note that the carefully-designed LaundryCo program adheres to two design principles that we will
address in the discussion section. First, participation was purely voluntary. While participants received a small
insurance premium reduction, non-participants’ rates did not change. LaundryCo management did not record
non-participants, and plant managers report not knowing the identity of non-participants. Second, LaundryCo
did not have access to the health data. We received these data directly from the vendor, and LaundryCo was
not party to any individual results from the program. All employees were informed that participation was
voluntary and that privacy was assured because LaundryCo would not administer the program.
The program began with a blood draw at the plants, which defines our treatment date. The blood samples
were then shipped to LabCorp, a national medical laboratory testing company, and tested for 42 common
health markers. Employees were also given a health survey from Wellsource, a provider of evidence-based
health assessments. The survey asked about health background, nutrition, fitness, stress level and mood, drug
use, and other behaviors. Because it was administered during work hours, and reviewed by the nurse, the
survey had few missing answers (2.2% of responses) that necessitated data exclusion. Approximately three
weeks after the screenings, nurses from the vendor held an educational seminar at each plant and presented
each worker with a personalized report detailing their results. About 97% of participants in our sample had at
least one abnormal test result each year, although some of these reflected small, likely-random variations on a
few blood test measures. Both LaundryCo’s HR director and interviews with workers indicated that because
many workers rarely visit doctors, many biometric screenings uncovered serious or unknown health

9"
"
problems.1 The nurse called approximately 20% of employees with severely abnormal results within a week
of the first visit, explained the results, then asked for permission to forward the results directly to the
employee’s physician. Employees who lacked a primary care physician were offered a referral. The health
packet detailed the individual’s health status, including blood results and abnormalities, health behavior
scores, anticipated future risks, and personalized suggestions for improvement. Even employees who were
aware of an existing medical condition, such as obesity or borderline diabetes, might not have known its
deleterious impact, or ways to remediate it, until they received this detailed, individualized information.
LaundryCo offered the full program, including biometric screenings, surveys, and counseling, to all
employees in 2010, 2011, and 2012. Although the program continues to operate for new hires, LaundryCo
stopped offering the full biometric screenings for previous participants starting in 2012.
3.2 Data
The data span 2009-2012 and cover all production workers at five LaundryCo plants for which productivity
data are available. Figure 2 illustrates the program structure with three sets of visits from the outside vendor.
LaundryCo offered the wellness program to all employees in 2010, 2011 and 2012. To measure whether an
employee’s health improves due to the wellness program, we limit our sample to employees who participated
in two screenings in a row, and limit our analysis to the first such instance (i.e., 2010 and 2011, or 2011 and
2012). The date of treatment is defined as the date of blood draw for the first instance of participation.
Turnover severely reduces our sample. Of the 347 production workers employed during our sample period,
56 stayed for less than a month, while an additional 106 stayed for only one instance of the program (98 in
the treatment plants and 8 in the control plant).
Our dataset combines three main sources. First, we received health data from the outside vendor. These
data include objective biometric blood tests and wellness survey results. Second, we employed outside
physicians to use the biometric and survey data to assess each employee’s health level and improvement
between years. Finally, LaundryCo provided demographic, human resource, and daily productivity data from
their IT system. We merged and de-identified these three datasets under an approved university IRB protocol.
<<< INSERT FIGURE 2 HERE >>>
3.2.1 Biometric data. The vendor compiled longitudinal blood data for each participant, including a
panel of 42 blood tests. These biometric screenings assessed abnormalities in diabetes, cholesterol, kidneys,
enzymes, iron, electrolytes, cell balance, thyroid, complete blood counts, white blood counts, and prostate
(PSA). The average worker had nearly five abnormal blood results out of 42 tests given. However, since the
1 There were many employees who explained how the wellness program helped improve awareness of their health. One employee
said, “They caught my thyroid! They let me know how out of balance I was and made referrals to help get me on the right
medication.” Another stated that, “Our wellness program helped me find out that I had diabetes…. Since the test I have been treated
and am feeling much better overall.” A third employee shared that, “During the blood draws last year it was discovered that I had
high potassium levels and also had a hyperactive thyroid…. Since then I have got with my physician and I am feeling much better.”
Finally, an employee shared this story: “[The wellness program] opened my eyes to something I did not know I had. I was unaware I
was diabetic or had high cholesterol. Since last year, I went to the doctor and have been put on medication and my diabetes and
cholesterol is [sic] under control. I have also lost 25 lbs. If this had gone unnoticed, I could be going down the same road as my
brother who has recently had to have toes removed due to his diabetes.”

10"
"
normal test ranges represent 95% confidence intervals, and since the tests themselves contain measurement
error, even healthy employees were likely to have at least a few abnormal results out of 42 tests. In addition,
the outside vendor measured blood pressure and calculated body mass index (BMI), a measure of obesity.
3.2.2 Survey results. The vendor also provided survey results from the Wellsource wellness
questionnaire, which participants completed at the time of testing and included 110 questions on health
history and behavior such as exercise and eating habits, drug use, sleep behaviors, current medications, mental
health, job satisfaction, health learning interests, and safety (e.g., seatbelt and sunscreen use). Figure A1 in the
Appendix shows the actual survey. The survey elements used in our analyses had only 2.2% of responses
missing, resulting in data exclusion.
3.2.3 Physician evaluation of biometric and survey data. We employed three experienced internal
medicine residents from a major Midwestern university hospital to thoroughly evaluate the pre-existing health
and health improvements of each employee. The doctors evaluated health and improvement first individually
and then as a group. We hired physicians because, while we have detailed health information on each
employee, there was no overall assessment of illness, or whether overall health improved between testing
events. Physician focus groups told us that it would be difficult to build objective measures of sickness based
on the data alone, and that doctors are trained to examine a mix of objective measures and subjective
evidence such as responses to questions during office visits. Based on this feedback, we chose to have a
group of physicians read the entirety of each participating employee’s file.
We asked each physician to individually answer four questions for each patient by evaluating their
complete biometric and survey records (see Figure A2 in the Appendix for questionnaire): (Q1) Do they likely
have one or more medical conditions? (Q2) How seriously ill is the patient (5-point scale)? (Q3) How much
would that problem impact their ability to carry out eight hours of manual labor (5-point scale)? (Q4) How
much did the employee’s health improve between annual tests (5-point scale)?
Inter-rater agreement (IRA) on Q1 was good, with Fleiss’ Kappa of 0.50. This means that for 90% of the
employees, all three doctors marked them as having a medical condition, or not having one. Inter-rater
reliability (IRR) levels on Q2-Q4, however, were much lower, with Krippendorff’s Alpha scores of 0.17, 0.12,
and 0.20. These values would be low for non-expert raters on simple tasks, but are consistent with many
studies of expert agreement and IRR in the medical literature (e.g., Einhorn 1972; Elmore et al. 1994; Beam et
al. 1996; Chiang et al. 2007). For example, a study assessing the diagnostic abilities of 32 medical interns by 12
full-time medical faculty using 128 different clinical evaluation exercise methods found IRRs ranging from
0.00 to 0.63, with a mean of only 0.23 (Kroboth et al. 1992). This means that medical faculty observing the
exact same interaction between an intern and a patient were in agreement about how successfully the intern
diagnosed a patient problem at close to the same rate as our physicians agreed on Q2-Q4, and at a much
lower rate than our physicians agreed on Q1. With results from so many blood tests as well as survey results,
nearly every patient-year observation contains some contradictory data, and physicians tend to focus on
different measures based on their own beliefs and experiences (Eddy and Clanton 1982). Furthermore, most

11"
"
studies of IRR in medicine involve a physical examination of the patient, which our physicians lacked.
Therefore, our low IRR ratings are consistent with the literature on the difficulty in medical evaluations.
In fact, the physicians themselves expressed to us that Q2, Q3 and Q4 were difficult given the large
amount of data and many borderline results, and after completing their individual ratings, they proposed
collective evaluations involving a scoring system established by the doctors, with higher values broadly
reflecting worse health (see Figure A3 in the Appendix for the scoring system). We use both the individual
and collective measures for Q4, and the results are highly similar. We chose not to use Q2 in the study for
several reasons. First, Q2 is simply a more detailed version of Q1, yet this detail produces poor agreement
across physicians and is therefore inferior to Q1. Second, the physicians indicated frustration with this
question and that the dichotomous choice of Q1 was more appropriate for general health assessment. We
excluded Q3 because the low IRR likely resulted from forcing physicians to match general health assessments
to job tasks they had not observed. In the case of both variables, the low IRRs indicate significant
measurement noise that would be prohibitive in a study that includes relatively few employees.
3.2.4
LaundryCo
IT data. LaundryCo provided data at the employee level on individual productivity and
demographics for 2009-2012. Productivity data are at the worker/day level, and the other data are measured
as of January, 2013. Demographic data include employee age, salary, tenure, and plant assignment—all of
which are absorbed in our fixed effect regressions. Worker productivity is based on how long an employee
works each day and how efficient they are on each given task.
3.3 Identification Strategy
We use a difference-in-differences (DiD) study design. The DiD design treats employees in LaundryCo’s four
plants that participated in the wellness program as the treatment group and the single non-participating plant
as a quasi-control group (hereafter called the control group). Because some treatment plant employees chose
not to participate in the program, we estimated a separate treatment effect for these employees. The DiD
strategy “differences out” fixed differences between treatment and control groups, and uses post-treatment
changes for the control group as a counterfactual for what would have happened had treatment group
individuals not participated in the wellness plan. The DiD approach is the most widely used methodology to
examine the impact of exogenous shocks or policy changes (Gertler et al. 2011).
Our treatment group is comprised of 69 employees who were present for at least two wellness program
exams at one of the four participating plants. We exclude from this group 81 employees who left employment
with LaundryCo after participating only once in the wellness program because our identification relies on
observing an employee’s health improvement in a second exam after their initial participation a year prior. We
furthermore drop 17 treatment group employees who opted out of participating in the wellness program a
single time, and who left the company before a second program could be administered. We also drop 67
treatment group employees who were not employed during a single evaluation. Our final treatment group
consists of two types of employees: the 55 workers who chose to participate and the 14 who did not, nine by

12"
"
choice and five because of absence on exam day.2 Following the literature on health intervention, we term
participating workers “compliers,” and the non-participants “non-compliers.” Both compliers and non-
compliers were employed at a treatment plant and received information about the program. Although non-
compliers were younger and had shorter work tenure, their average efficiency, pay, and daily hours worked
were not substantially different from compliers, as shown in Appendix Table A1. The control group is
comprised of workers from LaundryCo’s unionized fifth plant, which used a different insurance plan and did
not participate in the program. As with treatment group workers, we limit the control group to the 42
employees who were employed for the period encompassing at least two exams at treatment plants.
Although DiD strategies do not require identical treatment and control groups, in our case workers from
both groups were similar on most dimensions. All five plants work on the same tasks, use the same
production technology, and share common floor layouts. Three of the plants – the control plant and two
treatment plants – have approximately 30 employees working on a given day. The other two treatment plants
average 15 workers, but simply have fewer lines within a given category (e.g., one mat rolling station instead
of two). Also, two of the treatment plants are geographically proximate (32 and 34 miles) to the control plant,
which addresses local shocks such as weather that might affect productivity. Table A1 in the Appendix
presents pre-program demographics and efficiency for the three groups. As we note earlier, the one key
difference at the control plant is its unionized workforce and consequent lower turnover. Since our DiD with
worker fixed effects will absorb this fixed difference, this does not invalidate our identification strategy. If the
program itself led to changes in turnover our results would be biased; however, as seen in Appendix Figure
A4, there is no discernable impact on turnover.
Interviews with both corporate and control plant managers confirmed that no wellness program was
offered there. Interviews also revealed that nearly all other major plant policies were directed by the corporate
office, and affected all plants, not just the control or treatment plants. There was one exception – an
attendance awards program was put in place in the second year of our study at the control plant. We discuss
the implications of this program in section 4.2.3. Of course, managers may have their own management style,
but any fixed differences in management policies (or any other variable) do not affect DiD estimates.
DiD studies do not require random assignment of treatment and control, but a potential weakness of our
study is the lack of a true control group, as employees choose the plant at which to seek employment.
However, it is extremely unlikely that participants chose their plant based on expectations of a future wellness
program. Interviews suggest employees desire to work at the closest plant, and very few employees would
find it desirable to transfer to a different plant from their current location. Therefore, with respect to being
offered the wellness program opportunity, our intervention approaches random assignment.
3.4 Variables
2 Since the exams were known ahead of time by workers, employees may have chosen to be absent to avoid the exam, a
possibility acknowledged by management.

13"
"
3.4.1 Dependent variable. Our dependent variable is daily worker efficiency. LaundryCo uses a sophisticated
IT system that carefully tracks each worker’s productivity (called “efficiency”) on each task every day. To do
so, it measures the garment processing rate for each worker compared to the time-studied expected rate, as
determined by corporate headquarters. Scores are normalized such that 100 reflects performance that meets
expectations. For example, the time-studied rate for pressing dress shirts is 50.4 seconds, meaning an
employee must press over 71 shirts each hour to earn a score of 100. The system computes an overall daily
efficiency rate for each worker, equal to the weighted average (by time spent) of the worker’s efficiency scores
on each task that day. For example, a worker who spent two hours sorting soiled clothes with an efficiency
score of 80, two hours loading soiled clothes and the appropriate soap into washing machines with a score of
140, and four hours unloading clean but wet clothes into bins to be taken to the dryer area with a score of
160 would have a final daily efficiency of 135. A typical employee will have an efficiency number around 110-
120, with high performers consistently performing above 130 and low-performers at less than 100.
3.4.2 Independent Variables. Our independent variables represent four broad constructs: Compliers, Non-
compliers, Sick, and Better. Compliers takes a value of 1 for treatment plant employees who participated in the
program twice, and a 0 for non-complying and control group employees. Non-compliers takes a value of 1 for
employees who chose not to participate in the program when offered, or were absent, and a value of 0 for
compliers and control group employees. Empirically, these two variables are interacted with a dummy for the
post-treatment period, which for each employee is the date of initial wellness program participation.
In our primary health specifications, Sick is a dummy variable created from question 1 in the physician
health evaluation. Because IRR on this question is high, indicating 90% unanimity among the three doctors,
we require all three physicians to designate a worker as having a health condition based on the results of the
wellness intervention that year. Results were highly similar using a cutoff of two of three physicians
designating a worker as sick. There were 36 “sick” employees, and 19 “not sick” employees.
Similarly, in our main specifications, Better is a dummy variable using question 4 from the physician
evaluation that indicates significant health improvement between the first and second exam. We designate an
individual as Better if the average doctor score indicates health improvement between periods (i.e., average
>3). Fifteen workers qualified as “better,” and 40 workers were “not better.” Note that both “sick” and “not
sick” employees were classified as “better.” The largest of the four worker categories in Figure 1 was the 27
“sick, not better” employees. Thirteen workers were “not sick, not better,” six were “not sick, better” and 9
were “sick, better.” As stated previously, there were 14 total non-compliers at the treatment plant employed
during at least two blood draws.
We examined the use of continuous rather than binary measures of both Sick and Better, and while overall
model fit was superior, little additional insight is derived. We therefore present the results using the binary
measures, which are much easier to interpret. We present results using continuous measures in Appendix.
3.4.3. Control Variables. One of the main benefits of the DiD methodology is that fixed differences in
the control and treatment units do not affect the treatment estimate. Because fixed, pre-existing differences

14"
"
between groups are “differenced out” of the treatment estimate, including them in the regression (for
example as a time invariant dummy variable) will not influence the results. Additionally, as is standard when
using individual employee data, we include individual fixed effects, such that our treatment estimate
represents the change within an individual employee. Collectively, these factors mean that we do not require
an extensive set of control variables. Fixed differences across employees such as gender, ethnicity, and even
average capability are fully absorbed by individual fixed effects and cannot enter our specification.
However, employees do differ over time in their experience level; a given employee may be more efficient
several months after starting than they are in their first week. To control for within-employee learning, we
control for the cumulative months of job experience. The results are robust to using a dummy variable for
employees with less than a year of tenure, which interviews suggest is well after any learning-by-doing ceases.
Our last control is a plant-specific time trend; we use month fixed effects interacted with plant ID.
3.5 Specification
Our DID specification model is the following:
Yijt =
α
i +
β
1* Postt +
β
2*Postit*Non-Complieri +
β
3*Postit*Complieri +
β
4*Postit*Complieri*Betteri +
β
5*Postit*Complieri*Sicki +
β
6*Postit*Complieri*Sicki*Betteri +
λ
it +
γ
jt +
ε
ijt
where Yijt is efficiency for worker i in plant j, at time t, Posti is a dummy indicating all days after the first
screening for employee i; and Postit*Non-Complieri and Postit*Complieri are treatment group employees in the
post period who choose not to participate and to participate in the program, respectively.
λ
it is the
employee’s cumulative number of months of experience at LaundryCo, and
γ
jt is a plant specific time trend.
Our specification allows us to test for heterogeneous treatment effects depending on whether the program
identifies a worker as sick, and whether an employee’s health improves in the year after implementation. The
interactions between the Post variable (defined as 1 starting the date of the blood draw) and the Non-Complier
and Complier variables show the effect of the program on non-compliers and compliers, respectively, who are
neither sick nor better after the program. The subsequent triple and quadruple interaction coefficients
represent the incremental effect of the program on compliers who get better (
β
4), are sick (
β
5), and are both
sick and better (
β
6). We note that the time-invariant baseline effects, such as Non-Complieri, Complieri,
Complieri*Sicki, and Complieri*Betteri are all absorbed by employee fixed effects. We also note that we can
collapse complier categories and look at fewer complier groups by not differentiating by Sick, Better, or both.
All of the treatment coefficients (
β
2 to
β
6) are compared to the control group of workers not
participating in the program. Of course, we do not observe whether control group employees are sick or
improve their health. However, this does not bias our econometric results, because our sick and better variables
refer to sickness highlighted in the wellness program, and health improvements achieved due to the same
program. Control group workers by definition do not learn about their health or make health improvements
due to a wellness program, since they did not hear about or participate in any such program. Any changes to

15"
"
their health awareness, or their actual health, serve as the baseline counterfactual against which our wellness
program effects are measured. For example, general societal trends towards healthier eating and exercise, or
the effects of public policies such as the Affordable Care Act on health, are reflected in the control group and
are “differenced out” of the treatment estimate. The control group models the counterfactual of what would
have happened to treatment group employees absent the intervention.
We use ordinary least squares (OLS) to estimate our DiD model, clustering standard errors at the
individual level.
4. Results
As noted above, 65% of treatment group compliers (36 of 55) were classified as sick by the physicians, and
25% of sick workers improved their health after the intervention (9 of 36). Examples of common health
abnormalities were high cholesterol, obesity, hypertension, chronic pain, and self-admitted drug abuse. While
a 65% rate of sickness may seem high, the Center for Disease Control in the United States estimates that 50%
of adult Americans have at least one chronic health condition (Ward et al. 2014), although they note this is a
conservative estimate and does not include mental health, drug use, or obesity, which our doctors included.
Because of this broader definition of sickness, and because our sample is almost exclusively low-income
workers who have higher rates of illness than the general population, this higher sickness rate is not
surprising. Table 1 presents descriptive statistics for our main variables of interest, and Table 2 provides
correlations for the final sample. Note that Table 1 is at the worker-day level, which matches our regressions;
the statistics above on overall rates of sickness are at the worker level.
<<< INSERT TABLES 1 and 2 HERE >>>
Table 3 shows the formal statistical results from our regressions. Model (1) shows the overall effect of the
program on compliers and non-compliers; model (2) breaks down compliers into sick and not sick
employees; model (3) breaks down compliers into better and not better employees; and model (4) shows the
model with all four employee types. These models all use the independent doctor assessments to define the
Sick and Better variables; to be defined as Sick, all three doctors must judge a given employee to be sick based
on data from the initial blood draw. Appendix Table A2 shows the same models using the doctors’ collective
scale to define the Sick and Better variables, and the results are highly similar.
Rather than present the raw regression output, which shows the marginal effect of being in each
subsequent employee category (e.g., the marginal effect of being “better” on top of being “sick” and a
“complier”), Table 3 shows the total effect for each employee type, which represents the linear combination
of all relevant interaction coefficients. These should be interpreted as the average post-treatment productivity
change for each group relative to the control group. Marginal effects of model (4) are in Appendix Table A3.
The point estimate on model (1) suggests that the wellness program improved productivity among
compliers by 4.89 points, which is a 3.9% increase from an average productivity level of 125 points. Non-
compliers had reduced productivity of -7.21 points or 5.8%. Both estimates are statistically insignificant at

16"
"
conventional levels (p=0.229; p=0.179). However, models (2) and (3) both indicate that some employee types
saw statistically significant increases in productivity. In model (2), the productivity of non-sick employees
increases by 9.4 points or 7.5% (p=0.054), while model (3) estimates that the productivity of better employees
increases by 12.7 points or 10.2% (p=0.036). Results for those that do not improve are small and very
imprecise. In both models (2) and (3), the effect of the program on non-compliers is similar to model (1).
Finally, model (4) estimates separate treatment effects for the four groups in Figure 1. Because the number
of employees of each type is smaller than in models (2) and (3), the estimates are less precise, but the results
are fully consistent with the earlier models. Specifically, non-sick, better compliers (
β
=13.461, p=0.073) and
non-sick, non-better compliers (
β
=7.662, p=0.079) both see improved productivity. Sick, better compliers
also see better productivity, although the estimate is even less precise (
β
=11.778, p=0.106). Sick, non-better
compliers see almost no identifiable change in productivity (
β
=-3.068, p=0.468).
<<< INSERT TABLE 3 HERE >>>
The results from model (4) using all four employee groups are consistent with several of the mechanisms
that we presented earlier driving productivity improvements. First, improved job satisfaction or commitment
may explain why healthy compliers who do not get healthier still increase productivity. Second, even larger
productivity gains among both the healthy and sick who get better suggest improved capability through either
overall well-being or physical improvement. We cannot claim evidence for these mechanisms, but rather
present them as possible explanations for important individual productivity gains.
Our null effect for sick, non-improving workers casts doubt on the role of gratitude and reciprocity, but
we note several factors that may counteract gains among employees whose health does not improve. First, the
discovery of illness, and the failure to remediate it, may increase stress and depression, which have been
widely linked to decreased productivity and safety (Kuntz et al. 2014). Second, as we noted earlier, severe
health problems may not be easily addressable in the short-term. Although no employees were diagnosed
with a terminal illness, several had severe, uncontrolled diabetes, extremely high cholesterol, and morbid
obesity, problems might take more than a single year to rectify. While our results provide no evidence of
gratitude and reciprocity, employee interviews suggest these mechanisms were present in some cases.
4.1 Testing the Identifying Assumption of Parallel Trends
One of the main confounds of difference-in-differences models is that different pre-trends in the control
and treatment groups can generate spurious, significant results. The formal test for such pre-trends is to run a
“leads and lags” model that estimates separate “treatment effects” for a set of pre-treatment and post-
treatment periods (Autor 2003; Angrist and Pischke 2009). While the basic DiD model treats the entire pre-
treatment period as the baseline against which the post-treatment period is compared, the leads and lags
model uses a smaller pre-treatment period as the baseline. Other pre- and post-treatment periods are then
tested for treatment effects against this smaller baseline.

17"
"
Formally, this regression model is identical to the main regression specification outlined above but with a
series of time period dummies and their interaction with the treatment group. The omitted time period
provides the baseline difference against which each time interaction is tested. If pre-trends are driving the
effect, pre-treatment period coefficients would initially be large in the pre-treatment period and in the
opposite direction of the coefficient in the main model, with a marked trend towards the direction of the
model’s coefficient that begins well before the time of treatment. The lags and leads models also reveal if
treatment effects start immediately and if they persist.
Figure 3 shows quarterly leads and lags results for all five employee types in model (4) of Table 3: the four
complier types and non-compliers. The omitted quarter is the three months before program announcement,
since this is the closest period not influenced by the program. The first quarter after the omitted quarter
contains the program announcement, the blood draw, and the seminar, that occurred in consecutive months.
For the four complier types, there is no evidence of differential pre-trends—none of the pre-treatment
coefficients are large and negative in ways that might spuriously drive a positive coefficient in the baseline
model. None of the pre-treatment coefficients for these four types are statistically significant; the coefficients
are all less than 8 productivity points and almost always less than 4 points. Compared to the quarter just
before implementation, productivity differences between treatment and control groups are thus close to zero.
Furthermore, the largest pre-treatment coefficients (Figure 3C representing “not sick, better” employees) are
positive, which if anything biases the base DiD model towards a negative result, not a positive one.
<<< INSERT FIGURE 3 HERE >>>
For three of the four complier groups – not sick, not better; not sick, better; and sick, better – there is an
immediate increase in productivity differences between treatment and control groups after program
implementation, consistent with the program itself causing observed productivity improvements. For sick,
better employees, the productivity jump appears to be sustained throughout the five quarters after
implementation, suggesting that the health improvements brought about by the wellness program brought
sustained productivity increases. For not sick, better employees, the results are not as persistent; while the
point estimate remains positive throughout the sample, the effects decrease and become less precise starting
nine months after program implementation. We note that this cell of employees is the smallest in the sample
with only six employees. For not sick, not better employees, we see an immediate jump in productivity, but
this effect tapers off over time. This result suggests that productivity gains from the intervention may be short
lived if not accompanied by real improvements in health. Finally, for sick, not better employees, there appears
to be a short-term negative impact on productivity.
Plots of the raw data by group also suggest a treatment effect on some but not all workers, and that pre-
trends were not different across groups. Raw monthly efficiency averages are plotted for better compliers, not
better compliers and control employees in Appendix Figure A5. The version of the chart showing the average
pre- and post-treatment efficiency for these groups (as a horizontal line) suggests steady averages for not-
better compliers and control employees, but a large average increase for “better” employees. This chart is

18"
"
most similar to the DiD regression approach, for which no time trends are assumed, but we also show raw
data charts with linear and polynomial spline best fits. All three charts suggest a flat control group and an
unchanged “not better” complier group, with a large, discernable positive jump for “better” compliers.
The lags and leads results on non-compliers (Figure 3E) suggest their productivity was already declining
before program implementation. Differences between treatment and control group productivity compared to
the omitted quarter were fairly large, positive, and close to statistical significance at the five percent level. The
negative result on non-compliers in the statistical models is at least partly attributable to an apparent declining
trend in productivity of non-compliers that started well before the wellness program was implemented.
4.2 Robustness Tests
We conducted several robustness checks to these main findings. First, Appendix Table A4 shows
regression results using block-bootstrapped standard errors at the worker level (Cameron and Miller 2015),
with similar results. Second, we carried out a series of placebo tests, randomly assigning employees to
different groups, and randomly assigning intervention dates. This is to ensure that the serial correlation errors
that plague some DiD specifications are not generating spurious results (Bertrand et al. 2004). Figure 4 shows
placebo tests on the sick, better category of employees. For 47 of the 50 placebos run, the coefficient is both
smaller in size and less statistically significant than the estimated coefficient. Since the coefficient is significant
at the 10% level, this is approximately what one should expect from the placebo tests. Appendix Figure A6
shows placebo results on the other 3 complier groups, as well as non-compliers. We also ran robustness
checks on the non-complier category. In one such test, we excluded the five non-compliers who were absent
from the plant on the day of blood draw, as these employees might have not participated for exogenous
reasons such as sickness rather than by choice. These results are in Appendix Table A5. The results for non-
compliers is considerably larger and more significant when the five absent non-compliers are dropped;
however, the basic results for compliers do not change significantly.
Finally, we examined the impact of using continuous measures for our “sick” and “better” variables. Our
doctors provided 5-point scales for both variables, and we used the average of these raw scores in place of the
binary variables reported previously. However, there was not a lot of variation within these variables, and
interpreting the regression results and continuous interactions becomes very complicated, which is why we
used binary measures in the paper’s main results. In the Appendix, we show regression results using
continuous measures for these two variables in Table A6. Appendix Figure A7 shows a graphical
representation of the results, which confirms that employees who are the sickest and whose health most
improves see the largest productivity gains.
<<< INSERT FIGURE 4 HERE >>>
4.3 Specific Health Improvement Mechanisms
We next examine specific employee actions driving the health and productivity improvements. For these
tests, we primarily use survey data to identify workers who improve self-reported behaviors on nutrition,
exercise, and stress. We use the survey data to redefine our “Sick” and “Better” employee categories. We

19"
"
classify individuals as “Sick” if their score is below the minimum vendor-communicated threshold for that
dimension. We classify individuals as “Better” in each area if they improve their self-reported survey scores.
We additionally use blood data on HDL cholesterol to measure nutrition and exercise, as it is directly linked
to diet and exercise, is difficult to improve by medication alone (Wood et al 1988; Mensink et al 2003), and is
an objective rather than a self-reported measure. We note, therefore, that both the “sick” and “better”
variables were different from the doctor surveys in this set of results. Regression results using these new
measures suggest that productivity gains are driven by lifestyle improvements in exercise, nutrition, and stress.
Figure 5 shows that those who are “sick” on these dimensions, but who subsequently improve, see positive
productivity gains. All of these coefficients except for “nutrition” are significant at the p<.10 level.
<<< INSERT FIGURE 5 HERE >>>
Figure A8 of the Appendix repeats the analysis from Figure 5, with employees grouped by the original
“Sick” and “Better” variables defined by the three physicians. Although the number of employees in each
category is small, the point estimates indicate that lifestyle improvements by employees who were not sick
drive large productivity gains. These gains are correlated with lifestyle improvements in stress, exercise, and
HDL. It is apparent from these results that many employees without health problems made positive lifestyle
changes due to the program, and these changes drove significant productivity growth for LaundryCo. Overall
these results suggest capability improvements from improved overall and physical well-being.
We also examined whether improvements on specific health dimensions such as diabetes and kidney
function were identifiable as driving the productivity gains in our main results. Sick and Better were defined by
whether biometric test results were outside of or entered specified normal ranges, respectively. Although our
sample size weakens inference on these small subgroups, the results (Appendix Figure A9) are supportive of
the key mechanisms in Figure 5. The results show that improvements in disease areas associated with lifestyle
choices, such as diabetes, electrolytes, or cholesterol, led to large productivity growth for sick individuals,
although we caution that these data represent snapshots in blood tests that also may reflect random daily
variation in blood chemistry.
4.3 Addressing Empirical Limitations
4.3.1 Sample Size: Despite its strengths, our setting has two weaknesses that should be addressed in future
work. First, the number of workers for which we have two years of health data is only 111, which makes
precise estimation of productivity changes for different subsamples difficult. Our small sample partly reflects
LaundryCo’s size, but also is hindered by high turnover that limits the length of time over which we can
observe many workers. This problem is magnified by the small cell sizes for subgroup analysis. In addition to
the 42 employees at Plant 1 serving as a control group, 55 of the 69 employees at the treatment plants chose
to participate in the program. Of these 55, 36 qualified as sick and 15 improved their health. Only nine sick
and six non-sick workers improved their health. Furthermore, the problem is particularly acute when
attempting to identify the specific health improvements (e.g., diabetes) that drive productivity gains.

20"
"
Consequently, the imprecision of our coefficients should not be interpreted as strong evidence of a null
effect—we simply do not have enough statistical power to precisely identify small effect sizes.
Future work should seek organizations where larger samples will provide improved power and allow
possible identification of smaller effects that support (or refute) our findings. Larger organizations would also
allow for more detailed subsample analysis to pinpoint which specific health improvements are most crucial
to improving capability and productivity. That our study is the first to link objective productivity and health
data tells how difficult such a dataset would be to acquire. To help guide future research, we implement
simulated power tests for our own sample size and then project how larger samples might improve statistical
power. Power tests indicate the likelihood of a data sample correctly rejecting a null effect for a given
parameter magnitude at a specific significance level (a). To first test the power of our own sample, we
randomly generate 1000 datasets, each with 52,293 observations distributed across pre- and post- blood draw
periods for 111 workers: 42 are in the control group, 14 are non-compliers, 27 are sick non-improving
compliers, 9 are sick improving compliers, and 6 are non-sick improving compliers. The productivity values
for these observations are calculated using the parameter estimates in model (4) of Table 3 plus an error term
based on two components: worker-level and individual error terms. The two error terms are normally-
distributed random variables with mean zero and their respective standard deviations in our real sample, and
weighted based on the calculated intra-cluster (worker) correlation in our data (0.41).3
The statistical power calculations using our sample size of 111 workers are provided in Table 4 for p-
values of .05 and .10. As expected, statistical power is low, which helps explain the imprecision of our model
estimates. Clearly, both our analysis and future studies would benefit from significantly increased sample size,
although the non-existence of previous studies speaks to the difficulty of acquiring such samples.
<<< INSERT TABLE 4 HERE >>>
To provide guidance for future researchers, we repeat our simulation with different sample sizes, mapping
the relationship between sample size and statistical power given our real sample’s parameter estimates, cell-
size ratios, and error structure. These results, presented in Figure 6, show that firm samples with greater than
500 long-term employees are likely necessary to reliably estimate precise effect sizes similar to ours. Smaller
effects, such as may be the case for healthy, non-improving participants, would require even larger samples.
<<< INSERT FIGURE 6 HERE >>>
4.3.2 Addressing endogenous participation: The second weakness is the endogeneity of wellness program
participation within the four participation plants as well as the choice to actively try to improve health. It
could be that employees who agreed to participate were more likely to view the program in a positive light,
and more likely to commit to lifestyle changes. Indeed, recent work has shown that individual worker
differences such as time discounting or self-control can predict both health and other behavioral dimensions
(Gubler and Pierce 2014; Israel et al. 2014). Although these results suggest the potential for instruments to
3 These simulations were programmed in Stata with guidance from McConnell and Vera-Hernández (2015).

21"
"
address endogenous improvement, such instruments would put further pressure on the statistical power
issues discussed in the section above and are infeasible to implement here.
To address endogenous participation, we use local average treatment effects (LATE) models that estimate
the average causal effect on compliers by instrumenting the random assignment of intended treatment status
on endogenous compliance (Imbens and Angrist 1994). To implement this, we use a dummy variable
Post*TreatmentPlant as an instrument for Post*Complier in two-stage least squares models with individual fixed
effects that parallel models (1) and (2) in Table 3. The results, presented in Table A7 in the Appendix, show
unbiased parameter estimates that are consistent with our model, but expectedly less precise.
4.3.3 Attendance award program at control plant: The control plant started an attendance award program
in March 2011, nearly a full year into our study timeframe. Because it affected the control group only, its
introduction might bias the results of our study if the award program reduced productivity in the control
plant. Specifically, it might make the counterfactual time trend against which treatment group changes were
compared appear worse than they otherwise would have been. Alternatively, if the award program had a
positive effect on control group productivity, it would bias our study against finding an effect.
There are several reasons why the award program is unlikely to affect our results. First, research on the
award program found no overall effect on productivity, making it is unlikely that the award program at the
control plant is causing the results for the wellness program. The negative effect of the award program was
only for a small sub-sample of employees who had strong attendance behavior before the award was
implemented (Gubler et al. 2016). Second, if our results came from a deterioration in control group
productivity, we would not find heterogeneous treatment effects across treatment group categories, since all
control group members were assumed to be “Not Sick, Not Better” in our empirical design. That our empirical
results did find the heterogeneous treatment effects predicted by our study suggests that the results are driven
by treatment group changes, not control group changes. Indeed, raw plots of the data from Appendix Figure
A5 do not show a drop off in control group productivity, and do show positive jumps in productivity for
some (but not all) of the treatment subgroups.
5. Discussion and Conclusion
In this paper, we presented mechanisms through which firms might increase productivity by introducing
formal programs that help employees track and improve health and wellness. We explained why the
program’s effects on employee productivity might depend on both an employee’s pre-existing level of
sickness and their post-program improvements in health. Notably, we explained that firms should not simply
focus on enabling sick employees to identify and mitigate health concerns. Instead we argue that all types of
employees might improve their productivity after the introduction of a wellness program.
While our study did not examine the long-term persistence of these effects, we did examine a full year in
our lags and leads models. Although these models show that the large productivity gains from health
improvers are unlikely to be generated through a temporary Hawthorne effect, the smaller and less precise

22"
"
gains from non-improving participants appear to dissipate after about 9 months, as seen in Figure 3. Thus,
the motivational benefits from demonstrated commitment to workers may be short-lived, while the
capability-based benefits from health improvement may be more persistent. Still, we caution that some of the
persistent productivity gains we observe result from continued support by the firm for employee wellness. We
doubt that long-term gains would be achieved through a single or short-term intervention.
Our empirical setting has several unique characteristics that make it a natural laboratory for studying
employee health and productivity. First, our quasi-experimental setting and methods provide causal evidence
that builds on previous work that was almost all correlational. Second, our health improvement data include
both detailed objective medical tests as well as self-reported data. In addition, we used physician health
evaluations to determine true health improvements beyond the objective normal ranges for blood tests.
Third, our paper is the first to use objective productivity data, an important advance over biased self-reported
productivity measures. But as we detail above, the small worker sample size makes our estimates imprecise,
and they should be viewed as preliminary evidence. Future researchers would ideally study organizations with
thousands of participants in order to achieve the statistical power necessary to precisely identify effect sizes
similar to ours. We also note that such large samples would better allow researchers to dissect which specific
health improvements are crucial for productivity gains. Our results in Figure 5 suggest that diet and exercise
play key roles, but our small sample size limits our confidence in these conclusions.
5.1 Managerial Implications
Our empirical results demonstrate that the introduction of a corporate wellness program can have a large
impact on employee productivity, and therefore firm profitability. Yet our study also suggests that the ability
of an employer to enjoy a productivity-based return on investment (ROI) from the program crucially depends
on two factors: the participation rate and employee turnover.
To estimate the ROI of the program, we first worked with senior executives at LaundryCo to estimate the
value of a “free” hour of labor to the company; the estimate was that the labor itself was worth $15, while the
more efficient use of its fixed capital stock (i.e., spreading fixed costs over more productivity) was worth $9.
Therefore, an hour of “free” labor is worth $24 to the company. As seen in model (1) of Table 3, the average
complier saw an increase in productivity of about 4%, meaning they worked nearly an additional hour of
work per month due to the program. Assuming 220 days of work a year, and applying the coefficient estimate
on compliers, this increase in productivity is worth $1,690 per complier. Given that there were 55 compliers,
the benefit of the program for compliers was nearly $91,000. Non-compliers saw a productivity decrease of
nearly 6%; the same logic means that the non-complier productivity drop cost the company $2,528 per non-
complier. Since there were 14 non-compliers, the productivity decrease cost the company about $35,000. The
total productivity benefit is the difference between these two numbers, or $57,558. We do note that both the
estimates for compliers and non-compliers use imprecise coefficients from model (1), with p-values around

23"
"
0.2. Furthermore, to be conservative, we have accounted for the full productivity decrease of non-compliers,
even though the results suggest their productivity was already declining before the program was announced.
We next calculated the cost of the program. LaundryCo paid $120 per employee to the vendor for the
program, and estimates that it spent another $120 per employee in terms of lost work time (to take the tests,
have the follow-up visits, etc.). The program therefore cost $240 per participating worker. LaundryCo paid for
134 workers to take part in the program, 81 of whom left before they could participate a second time. This
means that the total cost of the program was $32,640. The total ROI of the program was therefore 76.3%.
It is striking that the ROI on the program is large, despite the fact that turnover is so high and that
conservative assumptions were used in its calculations. In the lower bound of our ROI estimates, LaundryCo
is paying the full cost of the program for the large majority of participants, and is enjoying no productivity
benefit from these participants; still, the ROI in this scenario is over 75%. Notably, the ROI is still relatively
high even though slightly over 20% of workers choose not to participate, and suffer from a costly decrease in
productivity, at least part of which was likely unrelated to the wellness program. Our analysis allows us to
estimate the returns to a hypothetical “perfect” program in a company without turnover and where all
employees chose to participate. In this case, the ROI would be nearly 590%.
Our results suggest even larger productivity gains for those whose health improved due to the program.
Notably, there is no difference in the productivity growth for health-improving employees who were and
were not identified as “sick” by the program. Again, this suggests that the impact of the program is more
widespread than one might initially think, improving the capability of workers across the health spectrum.
Indeed, survey results indicate that the wellness program led to lifestyle changes for employees regardless of
sickness levels, likely generating capability improvements in both sick and healthy workers.
The results also suggest some caution, as one group of employees – those whom the program identified as
sick but whose health did not subsequently improve – did not exhibit the productivity gains seen by the other
three groups. Consequently, we see no evidence of gratitude-based reciprocity. Additionally, either they did
not enjoy improved job satisfaction, or else such gains were cancelled out by the negative informational shock
about and/or subsequent treatment for the indicated disease. Some employees may receive truly devastating
news through a corporate wellness program, such as the existence of a terminal condition. This is unlikely to
be the case in our empirical setting, where the most serious sicknesses uncovered by the program involved
long-term manageable health conditions such as severe diabetes, obesity, and pain, not terminal diseases.
Thus, we cannot conclude what combination of mechanisms might have generated this non-result.
5.2 Boundary Conditions
Although our empirical setting demonstrates how wellness programs can improve productivity, two
important program design elements define boundary conditions for such improvements. First, employee
participation cannot be compulsory or heavily coerced through social pressure or financial penalties that
might induce psychological reactance. Psychological reactance theory (Brehm 1966) argues that individuals
strongly react to external influence that they perceive to restrict their autonomy. Programs such as wellness

24"
"
initiatives that threaten worker autonomy might motivate employees to assert their autonomy either through
resisting the program or even through reduced productivity. Indeed, public health scholars argue that strong
incentives and requirements in wellness programs can produce negative effects through psychological
reactance (Dowd 2002), since employees view health and lifestyle choices as outside their work domain. The
program in our empirical setting was both voluntary and only weakly incentivized.
Second, employees must trust that the firm will respect the privacy of employee health data and not use it
for employment-related purposes. HIPAA regulations in the United States forbid firms from accessing
employee health and wellness data collected through group health plans. However, data from employer-run
wellness programs may be legally accessible. While firms cannot legally use these data for employment
decisions, and must formally separate program administration from other human resource functions,
employees may not trust the firm to observe this prohibition. Firms must not only observe these regulations,
but also communicate and demonstrate this compliance for credibility with employees. Employees who
mistrust the firm’s use of private health data might view the wellness program as violating a broader
psychological contract that governs their overall relationship and influences their individual day-to-day actions
(Rousseau 1990). This perceived abrogation by the employer of a part of this implicit contract could reduce
overall job motivation, satisfaction, and retention among those who strongly value health privacy.
The imprecise negative effect estimate on non-compliers could reflect perceptions from these select
employees that the program threatened either their autonomy or privacy, thereby reducing their future
motivation despite not participating. This highlights the important difference between the reality of the
program and its perception by workers. Even the best-designed wellness programs may threaten a small
group of employees, and the firm must actively convince these workers of the program’s merits and safety.
References
Adams G, Flynn F, Norton M (2012) The gifts we keep on giving: Documenting and destigmatizing the
regifting taboo. Psychological Science, 23(10): 1145-1150.
Agle BR, Mitchell RK, Sonnenfeld JA (1999) Who matters to CEOs? An investigation of stakeholder
attributes and salience, corporate performance, and CEO values. Academy of Management Journal, 42(5): 507–
525.
Aguilera RV, Rupp DE, Williams CA, Ganapathi J (2007) Putting the S back in corporate social
responsibility: A multilevel theory of social change in organizations. Academy of Management Review, 32(3):
836–863.
Armeli S, Eisenberger R, Fasolo P, Lynch P (1998) Perceived organizational support and police performance:
The moderating influence of socioemotional needs. Journal of Applied Psychology, 83(2): 288.
Angrist, J. D., & Pischke, J. S. (2009). Mostly Harmless Econometrics: An Empiricist's Companion. Princeton
University Press.
Autor D (2003) Outsourcing at will: The contribution of unjust dismissal doctrine to the growth of
employment outsourcing. Journal of Labor Economics 21(1): 1-42.
Baicker K, Cutler D, Song Z (2010) Workplace wellness programs can generate savings. Health Affairs, 29(2):
304-311.
Barnett ML, Salomon RM (2006) Beyond dichotomy: The curvilinear relationship between social
responsibility and financial performance. Strategic Management Journal, 27(11): 1101–1122.

25"
"
Bartlett MY, DeSteno D (2006) Gratitude and prosocial behavior helping when it costs you. Psychological
Science, 17(4): 319-325.
Beam CA, Layde PM, Sullivan DC (1996) Variability in the interpretation of screening mammograms by US
radiologists. Archives of Internal Medicine, 156: 209-213.
Becker GS (2007) Health as human capital: synthesis and extensions. Oxford Economic Papers, 59(3): 379–410.
Bernstein ES (2012) The transparency paradox a role for privacy in organizational learning and operational
control. Administrative Science Quarterly 57(2): 181-216.
Berry L, Mirabito AM, Baun WB (2010) What's the hard return on employee wellness programs? Harvard
Business Review, December: 2012-68.
Bertrand M, Duflo E and Mullainathan S. How much should we trust difference-in-differences estimates?
Quarterly Journal of Economics (2004) 119(1): 249-275.
Beshears J, Milkman K, Schwartzstein J (2016) Beyond beta-delta: The emerging economics of personal
plans. The American Economic Review P&P, 106(5), 430-434.
Boles M, Pelletier B, Lynch W (2004) The relationship between health risks and work productivity. Journal of
Occupational and Environmental Medicine, 46(7): 737-745.
Brehm JW (1966) A Theory of Psychological Reactance (Academic Press, New York).
Buell RW, Kim T, Tsay CJ (2017) Creating reciprocal value through operational transparency. Management
Science 63(6): 1673-1695.
Burbano V (2017a) Social responsibility messages and worker wage requirements: Field experimental
evidence from online labor marketplaces. Organization Science. Forthcoming.
Burbano V (2017b) Getting gig workers to do more by doing good: Field experimental evidence from online
platform labor marketplaces. Unpublished working paper.
Burton WN, Chen CY, Conti DJ, Schultz AB, Pransky G, Edington DE (2005) The association of health
risks with on-the-job productivity. Journal of Occupational and Environmental Medicine, 14(8): 767–777.
Calzolari G, Nardotto M (2017) Effective reminders. Management Science. Forthcoming
Cameron AC, Miller DL (2015) A practitioner’s guide to cluster-robust inference. Journal of Human Resources,
50(2): 317-372.
Carmeli A, Gilat G, Waldman DA (2007) The role of perceived organizational performance in organizational
identification, adjustment and job performance. Journal of Management Studies, 44(6): 972–992.
Carnahan S, Kryscynski D, Olson D (2017) How corporate social responsibility reduces employee turnover:
Evidence from attorneys before and after 9/11. Academy of Management Journal. Forthcoming.
Center for Disease Control (2014) Diabetes Public Health Resource. Available at:
http://www.cdc.gov/diabetes/statistics/incidence/fig1.htm
Chan TY, Li J, Pierce L (2014a) Compensation and peer effects in competing sales teams. Management
Science 60(8): 1965-1984.
Chan TY, Li J, Pierce L (2014b) Learning from peers: Knowledge transfer and sales force productivity
growth. Marketing Science 33(4): 463-484.
Chapman LS (2012) Meta-evaluation of worksite health promotion economic return studies: 2012 update.
American Journal of Health Promotion, 26(4): TAHP-1.
Chatterji AK, Levine DI, Toffel MW (2009) How well do social ratings actually measure corporate social
responsibility? Journal of Economics & Management Strategy, 18: 125-169.
Chiang MF, Jiang L, Gelman R, Du YE, Flynn JT (2007) Interexpert agreement of plus disease diagnosis in
retinopathy of prematurity. Archives Ophthalmology, 125(7): 875-880.
Christian MS, Eisenkraft N, Kapadia C (2015) Dynamic associations among somatic complaints, human
energy, and discretionary behaviors experiences with pain fluctuations at work. Administrative Science
Quarterly, 60(1): 66-102.
Conti G, Heckman J, Urzua S (2010) The education-health gradient. American Economic Review, 100(2): 234–
238.
Currie J, Madrian BC (1999) Health, health insurance and the labor market: Chapter 50. In Orley C.
Ashenfelter and David Card, ed. Handbook of Labor Economics. Elsevier, 3309–3416.

26"
"
Dabos GE, Rousseau DM (2004) Mutuality and reciprocity in the psychological contracts of employees and
employers. Journal of Applied Psychology, 89(1): 52.
Dai H, Milkman K, Beshears J, Choi J, Laibson D, Madrian B (2012). Planning prompts as a means of
increasing rates of immunization and preventive screening. Public Policy & Aging Report 22(4): 16-19.
Dai H, Milkman KL, Hofmann DA, Staats BR (2015) The impact of time at work and time off from work on
rule compliance: The case of hand hygiene in health care. Journal of Applied Psychology 100(3): 846.
Danna K, Griffin RW (1999) Health and well-being in the workplace: A review and synthesis of the literature.
Journal of Management, 25(3): 357-384.
Dowd E (2002) Psychological reactance in health education and promotion. Health Education Journal 61(2):
113-124.
Eddy D, Clanton C (1992) The art of diagnosis: solving the clinicopathological exercise. New England Journal of
Medicine,
306(21): 1263-1268.
Einhorn HJ (1972) Expert measurement and mechanical combination. Organizational Behavior and Human
Performance, 7:86-106.
Eisenberger R, Armeli S, Rexwinkel B, Lynch PD, Rhoades L (2001) Reciprocation of perceived
organizational support. Journal of Applied Psychology, 86(1): 42.
Elmore JG, Wells CK, Lee CH, Howard DH, Feinstein AR (1994) Variability in radiologists’ interpretations
of mammograms. New England Journal of Medicine, 331(22): 1493-1499.
Flammer C (2015) Does corporate social responsibility lead to superior financial performance? A regression
discontinuity approach. Management Science, 61(11): 2549-2568.
Flammer C, Luo J (2017) Corporate social responsibility as an employee governance tool: Evidence from a
quasi-experiment. Strategic Management Journal 38(2): 163-183.
Gallup-Healthways (2012) Well-being Index. Available at:
http://www.healthways.com/solution/default.aspx?id=1125 [accessed December 9, 2012]
Gertler PJ, Martinez S, Premand P, Rawlings LB, Varmeersch CMJ (2011) Impact evaluation in practice (World
Bank Publications).
Glavas A, Piderit SK (2009) How does doing good matter? Effects of corporate citizenship on employees.
Journal of Corporate Citizenship, 36: 51–70.
Goetzel RZ, Hawkins K, Ozminkowski RJ, Wang S (2003) The health and productivity cost burden of the
“Top 10” physical and mental health conditions affecting six large U.S. employers in 1999. Journal of
Occupational and Environmental Medicine, 45(1): 5–14.
Goldstein SM (2003) Employee development: an examination of service strategy in a high-contact service
environment. Production and Operations Management 12(2): 186-203.
Grant AM, Gino F (2010) A little thanks goes a long way: Explaining why gratitude expressions motivate
prosocial behavior. Journal of Personality and Social Psychology, 98(6): 946.
Grant AM, Christianson M, Price R (2007) Happiness, health, or relationships? Managerial practices and
employee well-being tradeoffs. Academy of Management Perspectives 21(3): 51-63.
Gubler T, Larkin I, Pierce L (2016) Motivational spillovers from awards: Crowding out in a multitasking
environment. Organization Science 27(2): 286-303.
Gubler T, Pierce L (2014) Healthy, wealthy, and wise: Retirement planning predicts employee health
improvements. Psychological Science, 25(9): 1822-1830.
Hekman DR, Bigley GA, Steensma HK, Hereford JF (2009) Combined effects of organizational and
professional identification on the reciprocity dynamic for professional employees. Academy of Management
Journal, 52(3): 506-526.
Huckman RS, Pisano GP (2006). The firm specificity of individual performance. Evidence from cardiac
surgery. Management Science 52(4): 473-488.
Huckman RS, Staats BR (2011). Fluid tasks and fluid teams: The impact of diversity in experience and team
familiarity on team performance. Manufacturing & Service Operations Management 13(3): 310-328.
Imbens GW, Angrist JD (1994) Identification and estimation of local average treatment
effects. Econometrica, 62(2): 467-475.

27"
"
Ingraham C (2016) Nearly half of America’s overweight people don’t realize they’re overweight. Washington
Post. December 1, 2016.
Israel S, Caspi A, Belsky D, Harrington H, Hogan S, Houts R, Ramrakha S, Sanders S, Poulton R, Moffitt T
(2014) Credit scores, cardiovascular disease risk, and human capital. Proceedings of the National Academy of
Sciences, 111(48): 17087-17092.
Jones DA (2010) Does serving the community also serve the company? Using organizational identification
and social exchange theories to understand employee responses to a volunteerism programme. Journal of
Occupational and Organizational Psychology, 83(4): 857–878.
Kaiser/HRET (2012) Survey of Employer-Sponsored Health Benefits, 1999-2009. Available at:
http://www.hret.org/reform/projects/employer-health-benefits-annual-survey.shtml
KC DS (2013) Does multitasking improve performance? Evidence from the emergency
department. Manufacturing & Service Operations Management 16(2): 168-183.
Kitzmueller M, Shimshack J (2012) Economic perspectives on corporate social responsibility. Journal of
Economic Literature, 50(1): 51–84.
Kroboth F, Hanusa B, Parker S, Coulehan J, Kapoor W, Brown F, Karpf M, Levey S (1992) The inter-rater
reliability and internal consistency of a clinical evaluation exercise. Journal of General Internal Medicine
7(2):
174-179.
Kuntz L, Mennicken R, Scholtes S (2014) Stress on the ward: Evidence of safety tipping points in
hospitals. Management Science 61(4): 754-771.
Larkin, I (2011) Paying 30,000 for a gold star: An empirical investigation into the value of peer recognition to
software salespeople, working paper, Harvard Business School.
Lee JJ, Gino F, Staats BR (2014) Rainmakers: Why bad weather means good productivity. Journal of Applied
Psychology, 99(3): 504.
Margolis JD, Walsh JP (2003) Misery loves companies: rethinking social initiatives by business. Administrative
Science Quarterly, 48: 268-305.
McConnell B, Vera-Hernandez M (2015) Going beyond simple sample size calculations: A practitioner's
guide. Institute for Fiscal Studies Working Paper. (No. W15/17).
Medical Billing and Coding (2012) Available at: http://www.medicalbillingandcoding.org/blog/12-
companies-with-seriously-impressive-corporate-wellness-programs/
Mensink RP, Zock PL, Kester ADM, Katan MB (2003) Effects of dietary fatty acids and carbohydrates on
the ratio of serum total to HDL cholesterol and on serum lipids and apolipoproteins: a meta-analysis of 60
controlled trials. American Journal of Clinical Nutrition, 77(5): 1146-1155.
Milkman K, Beshears J, Choi J, Laibson D., Madrian B (2011) Using implementation intentions prompts to
enhance influenza vaccination rates. Proceedings of the National Academy of Sciences, 108(26): 10415-10420.
NCHS (2012) NCHS Data Brief No. 82. Available at: http://www.cdc.gov/nchs/data/databriefs/db82.pdf
Neumann W, Dul J (2010) Human factors: spanning the gap between OM and HRM. International Journal of
Operations & Production Management 30(9): 923-950.
Ødegaard F, Roos P (2014) Measuring the Contribution of Workers' Health and Psychosocial Work-
Environment on Production Efficiency. Production and Operations Management 23(12): 2191-2208.
Orlitzky M, Schmidt FL, Rynes SL (2003) Corporate social and financial performance: A meta-analysis.
Organization Studies, 24: 403-441.
Parks KM, Steelman LA (2008) Organizational wellness programs: A meta-analysis. Journal of Occupational
Health Psychology, 13(1): 58.
Podsakoff PM, MacKenzie SB, Lee JY, Podsakoff NP (2003) Common method biases in behavioral research:
a critical review of the literature and recommended remedies. Journal of Applied Psychology, 88(5): 879.
Rogers T, Milkman KL, John LK, Norton MI (2015) Beyond good intentions: Prompting people to make
plans improves follow-through on important tasks. Behavioral Science & Policy, 1(2): 33-41.
Rousseau DM (1990) New hire perceptions of their own and their employer’s obligations: A study of
psychological contracts. Journal of Organizational Behavior, 11: 389-400.
Song H, Tucker A, Murrell K, Vinson D (2017) Public relative performance feedback in complex service
systems: Improving productivity through the adoption of best practices. Management Science. Forthcoming

28"
"
Staats B, Dai H, Hofmann D, Milkman K (2016) Motivating process compliance through individual
electronic monitoring: An empirical examination of hand hygiene in healthcare. Management Science.
Forthcoming
Staats B, Gino F (2012) Specialization and variety in repetitive tasks: Evidence from a Japanese
bank. Management Science 58(6): 1141-1159.
Stewart WF, Ricci JA, Chee E, Morganstein D (2003) Lost productive work time costs from health
conditions in the United States: Results from the American productivity audit. Journal of Occupational and
Environmental Medicine, 45(12): 1234-1246.
Tan TF, Netessine S (2014) When does the devil make work? An empirical study of the impact of workload
on worker productivity. Management Science 60(6): 1574-1593.
Tan TF, Netessine S (2015) When You Work with a Super Man, Will You Also Fly? An Empirical Study of
the Impact of Coworkers on Performance. Working Paper.
Thayer RE, Newman JR, McClain TM (1994) Self-regulation of mood: strategies for changing a bad mood,
raising energy, and reducing tension. Journal of Personality and Social Psychology, 67(5): 910.
Ton Z. (2014). The good jobs strategy: How the smartest companies invest in employees to lower costs and boost profits.
Houghton Mifflin Harcourt.
Tsang JA (2006) Gratitude and prosocial behaviour: An experimental test of gratitude. Cognition & Emotion,
20(1): 138-148.
Turban DB, Greening DW (1997) Corporate social performance and organizational attractiveness to
prospective employees. Academy of Management Journal, 40(3): 658–672.
Vallgårda S (2012) Nudge: A new and better way to improve health? Health Policy, 104(2):200-203.
Wang H, Qian C (2011) Corporate philanthropy and corporate financial performance: The roles of
stakeholder response and political access. Academy of Management Journal. 54(6):1159–1181.
Ward, B, Schiller J and Goodman R (2014) Multiple chronic conditions among US adults: A 2012 update.
Preventing Chronic Disease. 11(April): 1-4.
Weber L (2014) Wellness programs get a health check, The Wall Street Journal, Available at:
http://www.wsj.com/articles/wellness-programs-get-a-health-check-1412725776
Wood PD, Stefanick ML, Dreon DM, Frey-Hewitt B, Garay SC, Williams PT, Superko HR, Formann SP,
Albers JJ, Vranizan KM, Ellsworth NM, Terry RB, Haskell WL (1988) Changes in plasma lipids and
lipoproteins in overweight men during weight loss through dieting as compared with exercise. The New
England Journal of Medicine, 319(18): 1173-1179.
Yee RW, Yeung AC, Cheng TE (2008) The impact of employee satisfaction on quality and profitability in
high-contact service industries. Journal of Operations Management 26(5): 651-668.
Zoller HM (2004) Manufacturing health: Employee perspectives on problematic outcomes in a workplace
health promotion initiative. Western Journal of Communication 68(3): 278-301.

29"
"
Figure 1: Four Participant Types and Possible Mechanisms for Productivity
Improvement
Figure 2: Timing of Wellness Program
Note: Blood draw dates refer to when blood was drawn and surveys filled out at each plant. Seminar dates refer to when
health information was returned in a personalized packet to each participant. This typically occurred 3-4 weeks post blood
draw. We only use data from the first time participating for each worker. The 2010 pre-period begins in 2009.
2nd cohort)pre)period
~Apr%2010 ~May%2011 ~June%2012
Worker)Performance
Worker)Performance Worker)Performance
1st cohort)pre)period 1st cohort)post)period
2nd cohort)post)peri od

30"
"
Figure 3: Lead and Lags Results
A. Not sick, not better compliers B. Sick, not better compliers
C. Not sick, better compliers D. Sick, better compliers
E. Non-compliers
Note: Dots represent estimate of difference between treatment and control groups compared to difference in Bucket -1.
Bucket 0 contains the program announcement, blood draw and seminar. Bucket -1, the omitted interaction, contains the
three months just before program announcement.

31"
"
Figure 4: Placebo Test on Sick, Better Employee Group
Note: 50 placebo simulations for the main model using randomized treatment groups and treatment dates.
Figure 5: Regression Coefficients Using Survey Measures
Note: Diamonds represent coefficient estimates for each specific category following the main specification.
Models are estimated using OLS with clustered standard errors at the individual level. “Sick” is defined as having
a nutrition score <50, hdl cholesterol <=39, not exercising, and reporting a stress signal. “Better” is defined as
positive improvements on a given category. Coefficient estimates are plotted for sick individuals that improve
(“Better” equal to 1). Estimates based off 14 improvers in stress, 12 in nutrition, 21 in exercise, and 37 in HDL.
-40 -20 0 20 40 60
Coefficient Estimate on Efficiency
010 20 30 40 50
Estimate Ranking
Placebo Coefficient Estimate Placebo 95% Confidence Interval
Treatment Coefficient Estimate
Random Treatment Groups

32"
"
Figure 6: Projected Statistical Power for Larger Worker Samples
Note: These two figures represent estimated statistical power for hypothetical different sample sizes for our main model
and its parameter estimates in Table 3. Sample sizes were increased in intervals of 10, from 51 workers to 1001 workers
while maintaining the same proportions of compliers, non-compliers, sick, and improving people. For each sample size, the
power estimate is based on 1000 simulated datasets. The vertical line at 111 workers represents our true sample.
Table 1: Descriptive Statistics at Worker/Day Level
Variable
Count
Mean
Sd
Min
Max
Efficiency
52293
125.83
31.60
24.67
264.10
Post period
52293
0.47
0.50
0
1
Month(continuous)
52293
17.64
10.05
0
42
Year
52293
2009
0.87
2009
2012
Participant
52293
0.52
0.50
0
1
Sick (doctor measure)
52293
0.33
0.47
0
1
Better (doctor measure)
52293
0.14
0.35
0
1
Better (doctor group measure)
52293
0.17
0.37
0
1
Better exercise days
52293
0.21
0.41
0
1
Better hdl
52293
0.35
0.48
0
1
Better nutrition score
52293
0.12
0.33
0
1
Better stress signals
52293
0.13
0.33
0
1
N
52293

33"
"
Table 2: Correlations for Primary Variables
Efficiency
Post
period
Participant
Sick
Better
Better
(group)
Better
exercise
Better
hdl
Better
nutrition
Better
stress
Efficiency
1.00
Post period
0.04
1.00
Participant
0.03
0.03
1.00
Sick
-0.13
0.04
0.67
1.00
Better
0.00
0.02
0.39
0.20
1.00
Better (group)
-0.03
0.03
0.43
0.29
0.91
1.00
Better exercise
0.08
-0.03
0.50
0.28
0.34
0.29
1.00
Better hdl
0.03
0.03
0.71
0.52
0.19
0.23
0.40
1.00
Better nutrition
-0.02
0.00
0.36
0.00
0.30
0.32
0.11
0.38
1.00
Better stress
-0.02
0.00
0.37
0.16
0.33
0.28
0.16
0.12
0.24
1.00
Table 3: Regression Estimates (Total Effects)
(1)
(2)
(3)
(4)
Dependent Variable
Efficiency
Efficiency
Efficiency
Efficiency
Model
OLS
OLS
OLS
OLS
Physician Scale
Independent
Doctors
Independent
Doctors
Independent
Doctors
Independent
Doctors
Non-compliers
-7.209
-7.128
-7.598
-7.456
(5.331) [0.179]
(5.244) [0.177]
(5.403) [0.162]
(5.328) [0.165]
Compliers
4.888
(3.971) [0.229]
Non-sick compliers
9.392*
(4.818) [0.054]
Sick compliers
0.979
(4.004) [0.807]
Non-better compliers
1.552
(3.657) [0.672]
Better compliers
12.670**
(5.969) [0.036]
Non-sick, non-better compliers
7.662*
(4.321) [0.079]
Sick, non-better compliers
-3.068
(4.213) [0.468]
Non-sick, better compliers
13.461*
(7.423) [0.073]
Sick, better compliers
11.778
(7.220) [0.106]
Control for worker experience
Y
Y
Y
Y
Plant time trends
Y
Y
Y
Y
Fixed effects
Y
Y
Y
Y
R2
0.455
0.458
0.463
0.465
# of employees
111
111
111
111
Observations
52293
52293
52293
52293
Sick cutoff
3 doctors
3 doctors
3 doctors
3 doctors
Note: Robust standard errors in parentheses, clustered by individual. P-values in brackets. The dependent variable is daily worker
efficiency. All three physicians must specify an individual as “sick” for “sick” to take the value of 1. Results are robust to a cutoff of 2
physicians specifying sickness. “Better” uses the average of the physician’s evaluations on improvement (Q4 in the Physician
Evaluation Questionnaire), and takes the value of 1 if the average indicates improvement (>3 on the Q4 5-point scale). ** p<.05
* p<0.10