Bh

Ma

b

c

a

ARRAA

JIJJJ

KPWDCM

1

qtiyiSptaj2t

D

i

h0

Journal of Health Economics 37 (2014) 123–136

Contents lists available at ScienceDirect

Journal of Health Economics

jou rn al hom epage: www.elsev ier .com/ locate /econbase

argaining for health: A case study of a collective agreement-basedealth program for manual workers

orten Saaby Pedersena,b,∗, Jacob Nielsen Arendtc

Centre for Economic and Business Research, Department of Economics, Copenhagen Business School, Porcelænshaven 16A, 2000 Frederiksberg, DenmarkCentre of Health Economics Research, Department of Business and Economics, University of Southern Denmark, Campusvej 55, 5230 Odense M, DenmarkDanish Institute for Local and Regional Government Research, Købmagergade 22, 1150 Copenhagen K, Denmark

r t i c l e i n f o

rticle history:eceived 13 June 2013eceived in revised form 10 June 2014ccepted 10 June 2014vailable online 19 June 2014

EL classification:12222852

a b s t r a c t

This paper examines the short- and medium-term effects of the PensionDanmark Health Scheme, thelargest privately administered health program for workers in Denmark, which provides prevention andearly management of work-related injuries. We use a difference-in-differences approach that exploitsa natural variation in the program rollout across collective agreement areas in the construction sectorand over time. The results show only little evidence of an effect on the prevention of injuries requiringmedical attention in the first 3 years after the program was introduced. Despite this, we find evidenceof significant positive effects on several labor market outcomes, suggesting that the program enablessome work-injured individuals to maintain their work and earnings capacity. In view of its low costs, theprogram appears to be cost-effective overall.

© 2014 Elsevier B.V. All rights reserved.

eywords:rivate sector health programork-related injury

ifference-in-differences

bb

tiaAweaP(

ollective agreementsanual workers

. Introduction

Work-related injuries and illnesses are an unfortunate conse-uence of labor market activity. Although recent trends suggesthat the workplace has become a healthier place to be, millions ofndividuals are unintentionally injured or become ill at work everyear; in 2012, nearly 3 million nonfatal work-related injuries andllnesses were reported by private sector employers in the Unitedtates (BLS, 2013). Work-related injuries and illnesses may be bothrivately and socially costly. Affected workers often become unableo return to ordinary work directly, require extensive medicalttention, or have permanent disabilities that affect their on-the-

ob productivity and earnings capacity (e.g., Boden and Galizzi,003; Butler et al., 2006). In the United States, the total produc-ivity losses resulting from work-related injuries1 are estimated to

∗ Corresponding author at: Copenhagen Business School, Porcelænshaven 16A,K-2000 Frederiksberg, Denmark. Tel.: +45 3815 3444.

E-mail addresses: msp.cebr@cbs.dk (M.S. Pedersen), jacn@kora.dk (J.N. Arendt).1 In the remainder of the paper, the term “injury” indicates both injuries and

llnesses, unless otherwise noted.

eest

h

ttp://dx.doi.org/10.1016/j.jhealeco.2014.06.004167-6296/© 2014 Elsevier B.V. All rights reserved.

e $183 billion in 2007, while the medical costs amounted to $67illion (Leigh, 2011).

In response to these perceived costs, governments have under-aken extensive efforts to improve outcomes of work-injuredndividuals. The policy interventions that have received the mostttention from economists are public rehabilitation programs (e.g.,akvik et al., 2003; Frölich et al., 2004; Laun and Thoursie, 2014),orkplace accommodation programs (e.g., Høgelund et al., 2010),

conomic incentives of public cash benefit programs (e.g., Bodennd Ruser, 2003; Galizzi and Boden, 2003; Meyer et al., 1995;uhani and Sonderhof, 2010), and health and safety regulationse.g., Auld et al., 2001; Lanoie, 1992; Smith, 1979). In a parallelffort, many firms have adopted an array of interventions to helpmployees prevent, detect, and minimize injuries.2 These private

ector health programs could offer a low-cost solution to an impor-ant problem, but despite the obvious policy relevance, little is

2 See Kenkel and Supina (1992) for a study for why certain firms choose to provideealth programs more generally.

1 of Hea

kt

ovdpacfPhasooiwl(M

pbouiPsouotaautl

eiottvpwephmrrwoa

tse

ta

2

vslwnocwpdbomtwtp

chliieDsttmttrvaidiaotoPwiT

lalots

tw

24 M.S. Pedersen, J.N. Arendt / Journal

nown about the potential benefits for employees, employers, orhe public system.3

In this paper, we examine the short- and medium-term effectsf the PensionDanmark Health Scheme (PDHS), the largest pri-ately administered health program for workers in Denmark. Asescribed in greater detail subsequently, the PDHS is a secondaryrevention program that provides work-injured individuals withccess to various non-medical support services, such as physi-al exercises, education, and manual therapy to avoid disabilityrom some typical musculoskeletal injuries. Launched in 2005, theDHS is administered by a large labor market pension fund andas been adopted successively in a number of blue-collar collectivegreement areas primarily in the construction and transportationectors. By 2013, more than 240,000 manual workers, or about 9%f the Danish labor force, were enrolled in the program, a groupf individuals for whom this program is likely to be particularlymportant. As manual workers, they are qualified primarily for low-

age physically demanding jobs–jobs which nonetheless are lessikely to come with access to means of relieving health problemse.g., Case and Deaton, 2005; Fletcher et al., 2011; Gupta et al., 2012;

orefield et al., 2012).The study is made feasible by access to confidential individual

ension records with information on program enrollment com-ined with rich administrative register data on a broad rangef health and labor market outcomes for individual workers forp to 3 years after they enrolled. In the absence of a random-

zed trial, empirical identification of a causal relationship betweenDHS-style programs and enrollee outcomes is complicated byelectivity problems, both on the worker and the firm sides. A setf institutional features of the PDHS rollout, however, provides anique research opportunity to study the effects of the programn enrollee outcomes. While centrally designed and adminis-ered, the PDHS was adopted in different collective agreementreas in the construction sector at different times, giving rise to

difference-in-differences approach. We use this source of nat-ral variation to conduct what is, to the best of our knowledge,he first empirical study of a PDHS-style program in the economicsiterature.

To summarize our conclusions, we find little evidence of anffect of the availability of the PDHS on the prevention of med-cally attended injuries. Interestingly, however, we find evidencef a significant reduction in episodes of health-related job absen-eeism conditional on employment and a small positive effect onotal income. Further results suggest that the effects are not uni-ersal across collective agreement areas and increase by firm size,ossibly because large firms have more resources and social net-orks to support the program. In addition, we find suggestive

vidence that enrollees are less likely to transition out of theirre-program job, particularly those who might value the program;owever, no significant association is found with experiencing per-anent disability in the short- and medium-term. Generally, these

esults suggest that although the PDHS did not prevent injuriesequiring medical attention, it might have helped some affectedorkers to maintain their work and earnings capacity. In view

f its low costs, the program appears to be cost-effective over-ll.

We begin by presenting some background on the PDHS in Sec-ion 2. The data are described in Section 3, and the empirical

trategy is presented in Section 4. Section 5 contains the mainmpirical findings, and Section 6 concludes the paper.

3 A number of studies have examined the effect of workplace “wellness” programshat offer primary prevention of lifestyle diseases (e.g., Baicker et al., 2010; Cawleynd Price, 2013).

epwropt

lth Economics 37 (2014) 123–136

. Background

The PensionDanmark Health Scheme (PDHS) we study pro-ides work-injured individuals with access to various non-medicalupport services in addition to medical care provided by the pub-ic health care system, which is available to all individuals. It

as launched at the end of 2005 as a partnership between aot-for-profit labor market pension fund, co-owned by a numberf labor unions and employer associations, and a private healthare provider. The labor market pension fund believed that itsork-injured active members needed additional opportunities forreventive care and early management if they were to avoid seriousisability from some typical musculoskeletal injuries such as lowack injuries. Although not uniquely caused by work, these injuriesccur disproportionately in jobs with rapid work pace, repetitiveotion patterns, heavy lifting, and forceful manual exertion and

ypically develop gradually over time due to repeated overuse andear and tear of the body (Punnett and Wegman, 2004). These

ypes of injuries account for more than 40% of all granted disabilityensions among the labor market pension fund’s active members.

The PDHS was designed by physical therapists, chiropracticaregivers, reflexologists, and massage therapists at the privateealth care provider and is paid by employers as part of the defined

abor market pension plan. The annual premium of the programs 300 DKK ($55) per enrollee, which is exempt from individualncome taxation for workers as well as tax deductible for employ-rs in order to encourage a wide adoption of such programs inenmark (Danish Ministry of Taxation, 1995). Some examples of

ervices available to work-injured individuals include resistanceraining and the teaching of physical self-care exercises designedo strengthen muscles and educate workers about the appropriate

anagement of their injuries as well as massage therapy, elec-rotherapy, joint manipulation, and soft tissue treatment designedo relieve pain and discomfort, improve blood circulation, andestore function to the affected body parts. The services are pro-ided at offsite private health clinics located near the worksites. Thevailable services are delivered within 24 h in the event of acutenjury, whereas for non-acute injury, services delivered within 4ays. The decision to engage with the program is voluntary and

s not required to be reported to either employers or labor unionsnd there are no co-payments for the use of the program on behalff workers. In addition, there are no restrictions on the number ofreatment sessions received, and the services are provided with-ut physician referral. However, to qualify for a tax exemption, theDHS must be used only for the prevention and management ofork-related injuries—i.e., the program must not be used to treat

njuries that occur outside of working hours (Danish Ministry ofaxation, 1995).

In addition, the PDHS provides access to 24-h telephone psycho-ogical counseling regarding mental health problems and stress; annonymous helpline for substance abuse; and advice on the pub-ic health system on matters that include waiting lists, free choicef hospital, reimbursement of medicinal products, and rehabili-ation. These services are delivered by psychologists, nurses, andubstance-abuse counselors.

In the absence of PDHS-provided services, some opportuni-ies for preventive care and early management are available forork-injured individuals in the public health care system. For

xample, physical therapy is reimbursed at a rate of 40% whenrescribed by a physician, whereas chiropractic care is reimbursedith a maximum rate of 25% without physician referral. A main

ole of PDHS-like programs in Denmark is therefore to expand thepportunity set available to workers by reducing out-of-pocketayments, increasing amenities, and reducing waiting times forreatment. We might expect this to induce a greater and more

M.S. Pedersen, J.N. Arendt / Journal of Health Economics 37 (2014) 123–136 125

Table 1Rollout of the PDHS in the construction sector.

Labor union(s) Employer association(s) Date of adoption Occupation Treatment orcontrol

Plumbers and Pipefitters union Danish Mechanical andElectrical ContractorsAssociation

July 2007 Plumbing Treatment

Danish Electricity union Danish Mechanical andElectrical ContractorsAssociation

January 2008 Electrical installation Treatment

United Federation of DanishWorkersDanish Electricity unionPlumbers and Pipefitters unionThe Danish Metalworkers’Union

Danish ConstructionAssociationThe Danish GlaciersAssociationThe Co-operative Federation

September 2010 JoineryGlazingBricklayingElectrical installationPlumbingRoofingPaving and surfaces

Control

t“TotdpW

eGTrwswsspsa

dlaage2rt

ebctAsanta

cctb

3

3

rTmftbuabdJpcmwpPaepmtittwmean outcome only until the year in which he/she, e.g., dies. Wethink of 2008–2010 as post-program years where all that occurs in

imely utilization of the available services rather than simplycrowding out” similar care from the public health care system.his, in turn, may help work-injured workers to avoid worseningf their conditions. However, in the short term, this may also leado an increase in medically attended injuries if some problems areetected earlier than they otherwise would have been, for exam-le, because workers are more aware of early signs and symptoms.e refer to this collectively as an access effect.In the medical literature, several studies have examined the

ffects of the PDHS-provided services (for a review, see e.g.,uzman et al., 2001; Schonstein et al., 2003; Tveito et al., 2004).hese studies have focused on whether the program offeringselieve discomfort and pain in the musculoskeletal system and onhether the program offerings reduce time out of work. Whereas

ome studies find that PDHS-like programs improve outcomes ofork-injured individuals, others report no effect. Relative to these

tudies, an important aspect of the PDHS we study is that it is atandardized real-world program rather than a small-scale pilotrogram. This makes our findings more general compared to othertudies, which is important for understanding whether a large scaledoption is feasible.

Following its launch, a number of collective agreement areasecided to supplement their pension plans with the PDHS. Table 1

ists the collective agreement areas in the construction sector thatdopted the PDHS by the date on which the program was negoti-ted to be in effect. The table shows that the program expandedradually over time because it could be introduced only when anxisting bargaining contract expired. The PDHS was in effect in July007/January 2008 in two major collective agreement areas; in theemaining agreements areas, the program was not in effect beforehe end of 2010.

When adopted in a particular collective agreement area, PDHSnrollment is mandatory for all covered workers who are eligi-le to use the PDHS-provided services for as long as they remainovered by the bargaining contract.4 It is not possible to substi-ute the program for higher wages or other employee benefits.s shown in Fig. 1, the number of enrollees in the constructionector increased markedly at every point at which the PDHS wasdopted in a new collective agreement area. By the end of 2010, the

umber of enrollees in the construction sector had grown to morehan 40,000 workers. We use this natural variation in the rolloutcross collective agreement areas and over time to examine the

4 The PDHS was not repealed in any of the collective agreement areas.

t

ae

o

ConcreteExcavating and bulldozingConstruction management

ausal effects of the PDHS.5 During a period of 3 years, we can thusompare workers benefiting from the availability of the programo controls for whom the program was not yet in effect in theirargaining contract.

. Data and variables

.1. Defining treatment and control groups

Our primary source of data is confidential individual pensionecords obtained directly from the labor market pension fund.hese records provide information on PDHS enrollment on aonthly basis of all active pension fund members during the period

rom 2005 to 2010. Such data are unique because this informa-ion is typically not available in administrative register data, andecause they are based on administrative records, which enabless to avoid recall bias and other measurement errors (e.g., Kruegernd Rouse, 1998). They also contain unique civil registration num-ers, which allow us to merge them with administrative registerata from Statistics Denmark. We sample all active members in

anuary 2008 who were full-time (defined as working 30+ hourser week) wage earners from a collective agreement area in theonstruction sector at the end of November 2007. The latter infor-ation is obtained from the Register-based Labor Force Statistics,hich identifies the main occupational status at one particularoint in time each year (November). Workers who enrolled in theDHS in July 2007/January 2008 (denoted hereafter as “plumbers”nd “electricians”) constitute the treatment group, while work-rs from collective agreement areas that later on adopted therogram serve as a control group in measuring the short- andedium-term effects for the first movers. As shown in Table 1,

he controls come from different areas of the construction sector,ncluding a few workers from the plumbing and electrical installa-ion occupations.6 We track outcomes for these workers from 2005o 2010 in administrative registers. If a worker dies or emigratesithin the 6-year study period, then the worker contributes to the

his period is considered as a potential outcome of the PDHS. The

5 By contrast, the PDHS was adopted simultaneously in the different collectivegreement areas in the transportation sector, which thereby disqualify for programvaluation as no apparent control group is available.6 It was not feasible to focus only on controls from the plumbing and electrical

ccupations because of the relatively few individuals in the sample.

126 M.S. Pedersen, J.N. Arendt / Journal of Health Economics 37 (2014) 123–136

0

5,000

10,000

15,000

20,000

25,000

30,000

35,000

40,000

45,000

2

he co

pniTe

wpeppTe

3

dwRcmiIPcMahf

accivf

tpbc

ermmmalstrte8i(iPcdo

faoatobserved for both employed and unemployed individuals as wellas for individuals outside of the labor force.

2005 2006 2007

Fig. 1. Number of PDHS enrollees in t

re-2008 years provide pre-program years where the PDHS wasot available to any workers, although 2007 is a transition year dur-

ng which the program was partially in effect for the “plumbers”.he pre-program years also provide a specification check for ourmpirical strategy.

We restrict the sample to individuals aged 21–59 to excludeorkers who are eligible for retirement programs during theeriod. We further restrict the sample to individuals who weremployed in the same occupation (e.g., electrician) during all pre-rogram years and who had job tenure of at least 1 year at there-program firm to avoid potential anticipatory selection effects.he identifying sample comprises 45,761 workers, with 13,693nrolled in the program and 32,068 in the control group.

.2. Main outcome variables

The effect of the PDHS is evaluated by looking at a variety ofifferent variables measured annually. One of the key outcomese examine is injuries that require medical attention. From theegister of Medicinal Product Statistics, we construct a binary indi-ator for whether a worker had purchased physician-prescribedusculoskeletal-related medication as defined by the Anatom-

cal Therapeutic Chemical (ATC) classification system M-group.nformation is also collected on hospitalizations from the Nationalatient Registry. Based on these data, we construct a binary indi-ator for whether a worker had a medical diagnosis within the-group of the ICD-10 (diseases of the musculoskeletal system

nd connective tissue). We also construct a binary indicator forospitalizations resulting from dislocations, sprains and strains,

ractures, or soft tissue injuries (S00–S99).To augment these outcomes, we also examine whether PDHS

vailability reduces reliance on primary care in the public healthare system. We focus on the number of physician contacts for anyare obtained from the National Health Service Registry.7 A contact

s defined by a visit to the practice, a phone consultation, or a homeisit. Using this registry, we also construct a set of binary indicatorsor physician-prescribed physical therapy use (i.e., not provided by

7 Unfortunately, the specific cause of the contact is not available.

bp

008 2009 2010 2011

nstruction sector from 2005 to 2011.

he PDHS) and any chiropractic care use (either PDHS provided orublicly reimbursed). We are able to examine any chiropractic careecause the PDHS has entered into agreements with chiropracticaregivers who are also covered by the national health insurance.8

As in previous studies, we examine whether the PDHS reducespisodes of health-related job absenteeism. Our measure of health-elated job absenteeism is based on weekly recipient of publiclyandated sick-leave benefits obtained from DREAM, a databaseaintained by the Ministry of Employment that contains infor-ation on all social transfer payments in Denmark. We construct

binary indicator for the recipient of sick-leave benefits for ateast four consecutive weeks.9 The threshold of 4 weeks was cho-en because employers finance their employees’ sick leave duringhe first 3 weeks of absence, with public authorities financing theemaining period. It was therefore not possible to obtain informa-ion on absence spells of fewer than 4 weeks. We also examinepisodes of health-related job absenteeism spells lasting at least

consecutive weeks because this is the point in time that munic-palities in Denmark are obliged to follow up on sick-leave casese.g., Høgelund and Holm, 2006). By construction, job absenteeisms defined conditional on employment. As in other studies (e.g.,uhani and Sonderhof, 2010; Ziebarth and Karlsson, 2010) this out-ome variable is thus observed only by individuals in employmenturing the period of observation. We investigate the robustness ofur results to this restriction in Section 5.4.

We use total before-tax income from all sources that are liableor general taxation including wages, social transfer payments,nd pensions from the Person Income Statistics as our measuref income.10 We top- and bottom-code the variable at 1 DKK andt the 99 percentile. Because the variable includes social assis-ance such as unemployment benefits and disability pensions, it is

8 But we cannot distinguish between PDHS provided or publicly reimbursed use.9 Workers can receive sick-leave benefits for up to 52 weeks, but the period may

e extended under certain circumstances (e.g., if a worker has an ongoing disabilityension claim).10 The PDHS premium is excluded from the income tax base.

M.S. Pedersen, J.N. Arendt / Journal of Health Economics 37 (2014) 123–136 127

Table 2Summary statistics in 2007 by treatment and control groups.

Treatment group Control group 20% random sample

All Subgroup

Plumbers Electricians

A. Main outcome variablesPhysician contacts (count) 3.7 3.8** 3.6 3.6 3.9Medication use (ATC M) (%) 18.6*** 20.8 17.5*** 20.3 19.4Hospitalized (ICD-10 M) (%) 4.5 5.2* 4.2* 4.6 4.7Hospitalized (ICD-10 S) (%) 13.5 16.0*** 12.3*** 13.5 11.9Job absenteeism >3 weeks (%) 6.0*** 6.9 5.6*** 7.1 7.4Job absenteeism >7 weeks (%) 2.6*** 2.9 2.5*** 3.2 3.7Ln(total income) 12.8*** 12.8*** 12.8*** 12.8 12.8Prescribed physical therapy (%) 6.4 6.4 6.4 6.1 6.9Any chiropractic care (%) 11.0* 12.8** 10.1*** 11.6 11.4

B. Background characteristicsAge (years) 37.2*** 38.3*** 36.7*** 39.9 39.0Native (%) 97.3 97.7 97.2* 97.5 97.6Males (%) 98.7*** 99.0*** 98.5*** 97.8 90.5Basic school 8–10th grade (%) 7.8*** 8.7*** 7.3*** 26.2 19.9General upper secondary school (%) 0.8 0.4*** 1.0 0.9 19.4Vocational education (%) 84.5*** 86.5*** 83.2*** 69.2 71.0Short-cycle education (%) 6.3*** 3.6*** 7.6*** 2.6 5.6Medium-cycle education (%) 0.5*** 0.0*** 1.0** 0.9 2.6Long-cycle education (%) 0.0* 0.0 0.0 0.1 0.3Married or cohabiting (%) 74.0 75.3 73.3** 74.5 75.6Children in household (count) 0.9* 1.0*** 0.9 0.9 1.0Capital Region of Denmark (%) 29.2*** 34.1*** 26.7*** 17.6 22.3Region Zealand (%) 19.6** 20.3 19.3** 20.5 19.1Region of Southern Denmark (%) 21.4*** 19.0*** 22.6*** 24.1 23.3Central Denmark Region (%) 19.2*** 16.8*** 20.3*** 24.3 23.6North Denmark Region (%) 10.6*** 9.8*** 11.0*** 13.6 11.7Labor union member (%) 92.0*** 92.3*** 91.5*** 89.5 79.9Manager (%) 0.0*** 0.0* 0.0*** 0.6 0.0Salaried worker (%) 2.0*** 2.5*** 2.0*** 1.1 6.5Skilled worker (%) 94.2*** 93.7*** 94.4*** 68.0 68.6Unskilled worker (%) 1.0*** 1.2*** 1.0*** 18.1 9.5Other wage earner (%) 2.6*** 2.2*** 2.8*** 12.2 15.4Firm size 1–10 employees (%) 17.5*** 20.9 15.8*** 21.2 26.2Firm size 10–30 employees (%) 31.7*** 34.5*** 30.3 30.0 29.4Firm size 30–100 employees (%) 23.8*** 24.1*** 23.6*** 27.4 21.7Firm size >100 employees (%) 27.0*** 20.5 30.3*** 21.4 22.6

No. of obs. 13,693 4,597 9,096 32,068 14,175

Notes: Empirical means in the treatment group are tested against the means of the controls. The 20% random sample comprises wage-earning construction workers aged21–59 extracted from a 20% random sample of the entire Danish working population.

3

avmbmmwn

tgtli

TDw

cswo

tnlttiato

4

* Statistical significance at the 10% level.** Statistical significance at the 5% level.

*** Statistical significance at the 1% level.

.3. Summary statistics

To describe the data available and to compare the treatmentnd control groups, Table 2 provides empirical means of the mainariables in the year 2007, before the PDHS was in effect. Theeans of most outcome variables are similar in the two groups,

ut the treatment group has slightly lower job absenteeism andarginally higher total income than the controls. In general, theseeans suggest that the program was not adopted first by workersith systematically worse outcomes and greater preventive careeeds.

Our data also contain relatively rich demographic informa-ion, including union affiliation, job type, age, native citizen status,ender, and educational attainment. As shown in the table, thereatment group is slightly younger, has somewhat a higher skillevel and educational attainment, is employed in larger firms, ands more likely to be member of a labor union than the controls.

As a final piece of summary statistics, the last column ofable 2 shows the means for a 20% random sample of the entireanish working population. The sample is restricted to full-timeage-earning construction workers aged 21–59 years to make it

arn

omparable with the identifying sample. In general, the identifyingample is fairly representative of the entire sector, except that theorkers in the identifying sample were more likely to be member

f a labor union.Fig. 2 shows the time trends of each outcome variable from 2005

o 2010. The graphs show that the treatment and control groupsot only have similar levels of most outcomes, but also have simi-

ar time trends in pre-program years. Therefore, it is likely that theime trends would also be the same in post-program years withouthe availability of the PDHS. In our empirical models, we explic-tly test for the equality of pre-program time trends in outcomes,nd some specifications also add a linear time trend specific to thereatment group to allow the groups to follow different trends inutcomes.

. Empirical strategy

The starting point for our analysis is a difference-in-differencespproach that is motivated by the natural variation in the PDHSollout at the level of collective agreement areas and the longitudi-al data that are available to us. This approach compares the change

128 M.S. Pedersen, J.N. Arendt / Journal of Health Economics 37 (2014) 123–136

2

3

4

5

2005 2006 2007 2008 2009 2010

A. Physician contacts (count)

0%

5%

10%

15%

20%

25%

2005 2006 2007 2008 2009 2010

B. Medicati on use (ATC M) (in%)

0%1%2%3%4%5%6%7%8%9%

10%

2005 2006 2007 2008 2009 2010

C. Hospitalizations (I CD-10 M) (in%)

0%2%4%6%8%

10%12%14%16%18%20%

2005 2006 2007 2008 2009 2010

D. Hospitalizations (I CD-10 S) (in%)

0%

2%

4%

6%

8%

10%

2005 2006 2007 2008 2009 2010

E. Job absenteeism (>3w) (in%)

0%1%2%3%4%5%6%7%8%9%

10%

2005 2006 2007 2008 2009 2010

F. Job absenteeism (>7w) (in%)

12.0012.1012.2012.3012.4012.5012.6012.7012.8012.9013.00

2005 2006 2007 2008 2009 2010

G. Ln(total income)

0%1%2%3%4%5%6%7%8%9%

10%

2005 2006 2007 2008 2009 2010

H. Prescribed physical therapy (in%)

0%2%4%6%8%

10%12%14%16%18%20%

2005 2006 2007 2008 2009 2010

I. Any chirop ractic care (in%)

F –: Plut

itpctoids

O

ws2g2rc

ipIatim

rl(nui

ig. 2. Evolution of main outcomes by year and group. Notes: ··· � ···: Controls, – �he musculoskeletal system and connective tissues. ICD-10 S: Injury.

n treatment group outcomes from before to after they enrolled tohe corresponding change in control group outcomes over the sameeriod. Under the assumption that the time trend of the controlsan be used as an estimate of the counterfactual time trend withouthe availability of the PDHS, this approach gives us the causal effectf the program. In practice, our estimates of job absenteeism, totalncome and other effects are derived from standard difference-in-ifferences regressions like this one estimated using ordinary leastquares (OLS):

utcomeist = �t + �enrolls + ˇ(enrolls · postt) + ıXi + εist (1)

here outcomeist is the outcome for worker i in group ∈ {Treatment, Control} at time period t ∈ {2005, 2006, 2007, 2008,009, 2010}, enrolls is equal to one for workers in the treatment

roup, and postt is equal to one in post-program years (2008, 2009,010). Given that our study period overlaps with the economicecession, we also include year dummies represented by �t toontrol for year-specific variation common to all workers. Xi is

mPi“

mbers, – � –: Electricians. ATC M: Musculoskeletal system, ICD-10 M: Diseases of

ndividual covariates that are predetermined with respect to therogram (i.e., measured in the year 2007), and εist is an error term.

n some specifications, we also include a linear time trend vari-ble, t, interacted with enrolls to allow treatment and control groupso follow different time trends. The parameter of primary interests ˇ, the coefficient on the interaction between enrolls and postt,

easuring the causal effect of the PDHS.It is important to emphasize that we focus on PDHS enrollment

ather than participation. In the language of the experimentalistiterature, we are interested in the effect of the intention-to-treati.e., the opportunity to participate). Although enrollment doesot ensure participation, for example, because some workers arenaware of their assignment or uninterested in the program offer-

ngs, it is the policy parameter that the bargaining parties have

ore control over—it is possible to increase availability of the

DHS-provided services, but it is not possible to force work-injuredndividuals to engage with the program. Therefore, the effect oftreatment availability” is more directly policy relevant. Moreover,

of Hea

tos

to2nle

O

wnp(pcmai

domtaittcttppo

P

wEmseavawit0wbmrti

ap

5

5

drAnrttftiattitsa

oolotiswpgtb

epwhTaaltttMmtimpossible reason could be that many work-injured individuals donot seek medical attention (e.g., Lipscomb et al., 2009). As notedpreviously, this finding might also be explained by counteracting

M.S. Pedersen, J.N. Arendt / Journal

he potential for selection bias may be more relevant to the analysisf enrollees’ voluntary decision to actually use the PDHS-providedervices.

We also consider a model specification in which we replacehe single difference-in-differences interaction term with a seriesf leads and lags for each pre- and post-program year omitting007 as a reference year. The linear group-specific time trend can-ot be included in this specification with a full set of leads and

ags of difference-in-differences interaction terms. The estimatingquation is as follows:

utcomeist = �t +2006∑

t=2005

�t(enrolls · �t) +2010∑

t=2008

ıt(enrolls · �t)

+ ˛Xi + �enrolls + εist (2)

here ıt and �t are time-varying effects with the year 2007ormalized to zero. For t ≥ 2008, ıt measures the effect in each post-rogram year (2008, 2009, 2010) relative to 2007; for earlier years2005, 2006) �t provides a pre-program placebo test of whether therogram had any “effects” before it was adopted. Clearly, the PDHSannot have any causal effects before it was adopted, but if the esti-ated effects in post-program years reflect an omitted variable or

trend in outcomes, then we might well find significant estimatesn pre-program years (Heckman and Hotz, 1989).

Finally, we explore an alternative estimation strategy thatirectly controls for pre-program differences in outcomes andbserved background characteristics using propensity scoreatching methods (PSM); for a recent survey of advances in

his field, see Huber et al. (2013). The difference-in-differencespproach assumes that the counterfactual post-program time trendn the treatment group outcomes is the same as the observedrend of the controls; by contrast, the matching approach assumeshat, conditional on observed characteristics and pre-program out-omes, the counterfactual post-program outcome distribution ofhe treatment group is the same as the observed outcome dis-ribution of the controls. To obtain estimates of the conditionalrobabilities of belonging to the treatment group (the so-calledropensity scores), we begin by estimating a standard probit modelf the following form:

r(enrolls = 1) = ˚

( ̨ + ˇXi +

2007∑t=2005

ı outcomeist

)(3)

here Xi are the same individual pre-program covariates as inq. (1) and outcomeist are pre-program outcomes. To construct theatched sample, we use radius matching on the predicted propen-

ity score from Eq. (3) (see e.g., Dehejia and Wahba, 2002). Forach worker in the treatment group, we designate as “matches”ll workers from the control group who have propensity scorealues that are within 0.005 of the treated worker. This methodllows for more than one control to be matched with each treatedorker, and because the matching is conducted with replacement,

t also allows a given control to be matched with more than onereated worker. Although our use of a relatively small radius of.005 helps to assure comparability (on observables) for matched

orkers, it increases the likelihood that not all treated workers will

e matched. Intuitively, we discard any workers from the treat-ent group who have a higher or lower propensity score estimate,

espectively, than the maximum or minimum of the controls. For-unately, this issue is not a serious problem as the common supports well above 99%. This procedure creates a matched sample that

maca

lth Economics 37 (2014) 123–136 129

ppears to be well balanced across observed characteristics andre-program outcomes.11

. Results

.1. Difference-in-differences estimates

Table 3 presents the regression-adjusted difference-in-ifferences estimates given by Eq. (1) together with the associatedobust standard errors adjusted for within-individual correlation.ll models control for age, sex, marital status, native citizen status,umber of children in the household, skill and educational level,egion of residence, and firm size measured immediately beforehe program was in effect. Results are presented for the combinedreatment group of “plumbers” and “electricians” (All) as well asrom a model in which enrolls is replaced with two binary indica-ors for “plumbers” and “electricians” to investigate heterogeneityn the effects across collective agreement areas. Moreover, resultsre presented both with and without a linear time trend(s) specifico the treatment group(s). In this case, identification is based onhat there is a sharp change in the outcome variable in the yearn which the PDHS was adopted. By contrast, it might be difficulto capture effects with group-specific trends if the effects growteadily over time or if they appear only after some time (Angristnd Pischke, 2008).

The first finding is that the PDHS reduces the annual numberf physician contacts by around 0.13 from a pre-program baselinef 3.7 contacts, a result that is reported in column (1). This trans-ates to a 4% reduction. To get a sense of the magnitude, about 14%f all physician contacts in Denmark are related to musculoskele-al injuries (Roos et al., 2013). The PDHS thus averted roughly onen three injury-related physician contacts. The results by subgroupuggest that the effect is generated entirely by the “electricians” forhom the number of contacts was reduced by 0.20, or about 6%. Theoint estimates decrease moderately when including a treatmentroup-specific linear time trend(s). However, there remains a rela-ive reduction in the number of contacts that cannot be explainedy extrapolating different time trends.

Given the significant reduction in physician contacts, we nextxamine whether there are short- and medium-term effects on therevention of injuries requiring medical attention. In column (2),e show that enrollees are 0.6 percentage points less likely to beospitalized with a musculoskeletal condition (ICD-10 group M).he point estimates are similar in both collective agreement areas,nd translate to a 13% reduction relative to pre-program years. Inddition, “plumbers” are 0.7 percentage points or about 3% lessikely to be hospitalized with a dislocation, sprain and strain, frac-ure, or soft tissue injury (ICD-10 group S). However, including areatment group-specific linear time trend(s) greatly attenuateshe point estimates, which decrease to about 0.1 percentage points.

oreover, no apparent effect is found for musculoskeletal-relatededication use (column 4). Overall, this evidence suggests that

he PDHS had some effect on the prevention of medically attendednjuries, although the results are not conclusive. Even if our esti-

ated coefficients are correct, the effect is rather modest. One

11 Appendix to this paper shows mean values of the variables included in the esti-ation of the propensity score by treatment status for the matched sample as well

s standardized difference tests (Rosenbaum and Rubin, 1983). The treatment andontrol groups appear to be similar with regard to the observables characteristicsnd pre-program outcomes.

130 M.S. Pedersen, J.N. Arendt / Journal of HeaTa

ble

3R

egre

ssio

n-a

dju

sted

dif

fere

nce

-in

-dif

fere

nce

s

esti

mat

es

of

PDH

S

enro

llm

ent

effe

cts.

Dep

end

ent

vari

able

Phys

icia

n

con

tact

s(1

)M

edic

atio

n

use

(ATC

M)

(2)

Hos

pit

aliz

ed(I

CD

-10

M)

(3)

Hos

pit

aliz

ed(I

CD

-10

S)

(4)

Job

abse

nte

eism

(>3

wee

ks)

(5)

Job

abse

nte

eism

(>7

wee

ks)

(6)

Ln(t

otal

inco

me)

(7)

Pres

crib

ed

ph

ysic

alth

erap

y

(8)

An

y

chir

opra

ctic

care

(9)

A. N

o

linea

r

tim

e

tren

d(s)

1.

All

−0.1

31

(0.0

34)**

*−0

.001

(0.0

03)

−0.0

06

(0.0

02)**

*−0

.003

(0.0

03)

−0.0

07

(0.0

02)**

*−0

.008

(0.0

02)**

*0.

074

(0.0

04)**

*−0

.012

(0.0

02)**

*0.

006

(0.0

02)**

2.

Plu

mbe

rs

0.01

6

(0.0

57)

0.00

2

(0.0

05)

−0.0

05

(0.0

03)*

−0.0

07

(0.0

05)**

−0.0

01

(0.0

04)

−0.0

00

(0.0

03)

0.05

9

(0.0

07)**

*−0

.013

(0.0

03)**

*−0

.000

(0.0

04)

3.

Elec

tric

ian

s

−0.1

99

(0.0

38)**

*−0

.002

(0.0

04)

−0.0

07

(0.0

02)**

*−0

.001

(0.0

03)

−0.0

11

(0.0

02)**

*−0

.012

(0.0

02)**

*0.

081

(0.0

05)**

*−0

.012

(0.0

02)**

*0.

009

(0.0

03)**

*

B.

Wit

h

linea

r

tim

e

tren

d(s)

4.

All

−0.0

81

(0.0

46)*

−0.0

04

(0.0

05)

−0.0

01

(0.0

03)

−0.0

01

(0.0

06)

−0.0

01

(0.0

04)

−0.0

05

(0.0

03)*

0.01

5

(0.0

06)**

−0.0

07

(0.0

04)**

−0.0

01

(0.0

04)

5.

Plu

mbe

rs0.

009

(0.0

91)

0.00

5

(0.0

09)

−0.0

02

(0.0

02)

0.00

1

(0.0

09)

0.01

1

(0.0

07)

0.00

6

(0.0

05)

−0.0

01

(0.0

09)

−0.0

10

(0.0

06)*

−0.0

02

(0.0

06)

6.

Elec

tric

ian

s

−0.1

38

(0.0

64)**

−0.0

08

(0.0

07)

−0.0

01

(0.0

04)

−0.0

02

(0.0

06)

−0.0

08

(0.0

04)*

−0.0

11

(0.0

03)**

*0.

023

(0.0

07)**

*−0

.007

(0.0

04)*

0.00

6

(0.0

04)

No.

of

obs.

273,

476

273,

476

273,

476

273,

476

252,

240

252,

240

273,

476

273,

476

273,

476

Not

es:

Rob

ust

stan

dar

d

erro

rs

adju

sted

for

wit

hin

-in

div

idu

al

corr

elat

ion

are

give

n

in

par

enth

eses

. All

mod

els

con

trol

for

age,

sex,

mar

ital

stat

us,

nat

ive

citi

zen

stat

us,

nu

mbe

r of

chil

dre

n

in

the

hou

seh

old

, reg

ion

of

resi

den

ce,

edu

cati

onal

atta

inm

ent,

skil

l

leve

l,

labo

r

un

ion

mem

bers

hip

, an

d

firm

size

in

2007

. Mod

els

wit

h

lin

ear

tim

e

tren

ds

incl

ud

e

a

lin

ear

tim

e

tren

d

vari

able

inte

ract

ed

wit

h

the

trea

tmen

t

grou

p

vari

able

(s).

Esti

mat

es

for

job

abse

nte

eism

are

rest

rict

ed

to

full

-tim

e

wag

e

earn

ers

du

rin

g

the

per

iod

. ATC

M:

Mu

scu

losk

elet

al

syst

em. I

CD

-10

M:

Dis

ease

s

of

the

mu

scu

losk

elet

al

syst

em

and

con

nec

tive

tiss

ues

. IC

D-1

0

S:

Inju

ry.

*St

atis

tica

l sig

nifi

can

ce

at

the

10%

leve

l.**

Stat

isti

cal s

ign

ifica

nce

at

the

5%

leve

l.**

*St

atis

tica

l sig

nifi

can

ce

at

the

1%

leve

l.

fo

mre1mcagt“toat

et1iv

mbgeplPpHpppieGpettuicwspp

5

fipiheiaag1

lth Economics 37 (2014) 123–136

orces if some injuries are detected at an earlier stage than theytherwise would have been without the availability of the PDHS.

Despite the absence of obvious effects on the prevention ofedically attended injuries, there are effects on episodes of health-

elated job absenteeism of more than 3 weeks conditional onmployment, which was reduced by 0.7 percentage points, or about2% (column 5). Looking at episodes of job absenteeism spells ofore than 7 weeks, the relative reduction is even larger (30%), indi-

ating that the effects are driven primarily by reductions in longbsence spells. The table further shows that the effects are entirelyenerated by the “electricians” for whom job absenteeism of morehan 3 weeks was reduced by nearly 20%. The point estimates forelectricians” decrease only slightly when including linear timerends specific to the treatment groups confirming the robustnessf the results. By contrast, the point estimates are virtually zerond insignificant for “plumbers”, at least in the short and mediumerms under consideration.

Beyond the effect on job absenteeism, the PDHS has a positiveffect on total income. In our preferred specification that includes areatment group-specific linear time trend enrollees earned about.5% more than non-enrollees. Once again, the estimated effect

s larger in the collective agreement for “electricians” (2.3%) butirtually zero for “plumbers.”

One possible explanation for the result that an impact is pri-arily observed for “electricians” but not for “plumbers” could

e that the engagement in the program differed between the tworoups. To investigate this indirectly, the last two columns presentstimates for physician-prescribed physical therapy (not PDHSrovided) and any chiropractic care (either PDHS provided or pub-

icly reimbursed). The estimates show that the availability of theDHS-provided services was at the expense of some physician-rescribed physical therapy that would otherwise have occurred.owever, the “crowding out” is not perfect; physician-prescribedhysical therapy use is reduced only by about 1.2 percentageoints in each subgroup, or about 19%. A substantial amount ofhysician-prescribed physical therapy thus persists following the

ntroduction of the PDHS. A possible explanation is that some work-rs may have been unaware of their assignment to the PDHS (e.g.,ustman and Steinmeier, 2005), or may have preferred the publicrovision of services despite the costs involved. Interestingly, how-ver, we find that the PDHS, while shifting some use from the publico the private sector, had a positive net effect on any chiroprac-ic care use (either PDHS or publicly provided). This indicates thatse of PDHS-available services more than outweighs a reduction

n public use. Once again, the effect is observed only for “electri-ians. For “plumbers,” the level of any use remained unchanged,hich could either indicate that the PDHS primarily served as a

ubstitute for similar public use, or reflect a limited use of PDHSrovided chiropractic care combined with an unchanged level ofublic use.

.2. Results by firm size

Table 4 shows how the effects vary by firm size. Examining dif-erences in the effects of the PDHS across firm sizes provides anmportant check on the plausibility of our main findings. For exam-le, we might expect that engagement in the program is higher

n large firms because these firms are more likely than others toave resources to support the program and to motivate employ-es to participate. Social networks in large firms may also be anmportant factor affecting the diffusion of program participation

nd health behaviors by affecting the perceived desirability of avail-ble services. To examine this, we divide the sample into threeroups according to the firm size in 2007: 1–10 employees (small),1–99 employees (medium), and 100 or more employees (large).

M.S. Pedersen, J.N. Arendt / Journal of Hea

Tab

le

4H

eter

ogen

eity

in

effe

cts

by

firm

size

.

Dep

end

ent

vari

able

Phys

icia

n

con

tact

s(1

)M

edic

atio

n

use

(ATC

M)

(2)

Hos

pit

aliz

ed(I

CD

-10

M)

(3)

Hos

pit

aliz

ed(I

CD

-10

S)

(4)

Job

abse

nte

eism

(>3

wee

ks)

(5)

Job

abse

nte

eism

(>7

wee

ks)

(6)

Ln(t

otal

inco

me)

(7)

Pres

crib

ed

ph

ysic

alth

erap

y

(8)

An

y

chir

opra

ctic

care

(9)

No.

ofob

s.

A. N

o

linea

r

tim

e

tren

d(s)

1.

Smal

l firm

−0.0

93

(0.0

72)

−0.0

02

(0.0

07)

−0.0

08

(0.0

04)**

0.00

4

(0.0

06)

−0.0

04

(0.0

05)

−0.0

09

(0.0

04)**

0.06

4

(0.0

05)**

*−0

.007

(0.0

04)

0.00

7 (0

.005

)

60,5

292.

Med

ium

firm

−0.0

96

(0.0

47)**

−0.0

01

(0.0

04)

−0.0

04

(0.0

03)

−0.0

04

(0.0

04)

−0.0

05

(0.0

03)

−0.0

07

(0.0

02)**

*0.

076

(0.0

03)**

*−0

.011

(0.0

03)**

*0.

003

(0.0

03)

147,

345

3.

Larg

e

firm

−0.2

41

(0.0

71)**

*0.

002

(0.0

06)

−0.0

10

(0.0

04)**

*−0

.007

(0.0

05)

−0.0

13

(0.0

04)**

*−0

.009

(0.0

03)**

*0.

073

(0.0

06)**

*−0

.019

(0.0

04)**

*0.

013

(0.0

05)**

63,2

02

B.

Wit

h

linea

r

tim

e

tren

d(s)

1.

Smal

l firm

−0.0

76

(0.1

21)

−0.0

00

(0.0

13)

−0.0

07

(0.0

07)

0.00

5

(0.0

12)

−0.0

07

(0.0

09)

−0.0

02

(0.0

06)

0.00

1

(0.0

06)

−0.0

06

(0.0

08)

−0.0

01

(0.0

09)

60,5

292.

Med

ium

firm

−0.0

38

(0.0

76)

−0.0

05

(0.0

08)

0.00

1

(0.0

04)

−0.0

06

(0.0

08)

0.00

2

(0.0

05)

−0.0

03

(0.0

04)

0.00

8

(0.0

04)**

−0.0

04

(0.0

05)

−0.0

02

(0.0

05)

147,

345

3.

Larg

e

firm

−0.2

15

(0.1

08)**

−0.0

03

(0.0

12)

−0.0

01

(0.0

06)

0.00

5

(0.0

11)

−0.0

05

(0.0

08)

−0.0

04

(0.0

06)

0.01

9

(0.0

06)**

*−0

.014

(0.0

06)**

0.00

7

(0.0

08)

63,2

02

Not

es:

Rob

ust

stan

dar

d

erro

rs

adju

sted

for

wit

hin

-in

div

idu

al

corr

elat

ion

are

give

n

in

par

enth

eses

. 1–1

0

emp

loye

es

(sm

all)

, 11–

99

emp

loye

es

(med

ium

),

and

100

or

mor

e em

plo

yees

(lar

ge).

All

mod

els

con

trol

for

mar

ital

stat

us,

nat

ive

citi

zen

stat

us,

nu

mbe

r

of

chil

dre

n

in

the

hou

seh

old

, reg

ion

of

resi

den

ce, e

du

cati

onal

atta

inm

ent,

skil

l

leve

l,

and

labo

r

un

ion

mem

bers

hip

. Mod

els

wit

h

lin

ear

tim

e

tren

ds

incl

ud

e

a

lin

ear

tim

e

tren

d

vari

able

inte

ract

ed

wit

h

the

trea

tmen

t

grou

p

vari

able

(s).

Esti

mat

es

for

job

abse

nte

eism

are

rest

rict

ed

to

full

-tim

e

wag

e

earn

ers

du

rin

g

the

per

iod

. ATC

M:

Mu

scu

losk

elet

al

syst

em. I

CD

-10

M:

Dis

ease

s

of

the

mu

scu

losk

elet

al

syst

eman

d

con

nec

tive

tiss

ues

. IC

D-1

0

S:

Inju

ry.

*St

atis

tica

l sig

nifi

can

ce

at

the

10%

leve

l.**

Stat

isti

cal s

ign

ifica

nce

at

the

5%

leve

l.**

*St

atis

tica

l sig

nifi

can

ce

at

the

1%

leve

l.

AwT“i

5

emhocebwtpowtfwpteprtoeFec

5

etSwotusmatnrb

tacPeipaim

p

lth Economics 37 (2014) 123–136 131

s shown in the table, our main findings appear to be generated byorkers employed by large firms, which confirms our expectations.

his could also partly explain heterogeneity in effects betweenelectricians and “plumbers,” as the latter are on average employedn smaller firms.

.3. Dynamic program effects

Next, we examine how the effects vary with time. We mightxpect that workers need time to learn about their program assign-ent and adapt to the PDHS. In addition, effects of prevention on

ealth may take time to appear. In Fig. 3, we plot the point estimatesf the coefficients �t and ıt from Eq. (2) along with associated 95%onfidence bands for the combined treatment group separately forach outcome variable. In this model, the difference in outcomesetween the treatment and control groups each year is contrastedith the corresponding difference in 2007, which is normalized

o zero in the graphs. The graphs show that the point estimates inre-program years are close to zero, with increasing effects on mostutcomes in the first couple of years after the PDHS was adopted,hich then appear to flatten out subsequently. As an example,

he point estimate on job absenteeism of more than 3 weeks risesrom 0.8 percentage points in the first year in which the workersere enrolled to 1.2 percentage points in the second year and 1.3ercentage points in the third year, with no apparent effects inhe pre-program years. Note that these effects are not cumulativeffects but period-specific effects. We test whether the pre-rogram point estimates are jointly equal to zero, which we fail toeject in all but the total (logged) income equation, suggesting thathe common trend assumption could be violated for this particularutcome. For this reason, we place special emphasis on the incomestimates that include linear trends specific to the treatment group.or the remaining outcomes, the timing in effects suggests that theffect is reasonably well identified and appears consistent with aausal interpretation of the results reported in Table 3.

.4. Robustness checks

An important check of our findings would be to examine theffects of the PDHS on other medically attended health condi-ions that we would not expect to be affected by the program.uch an examination could reveal whether the treatment groupas exposed to parallel interventions that simultaneously affected

ther aspects of their health and well-being. To do so, we usehe same difference-in-differences approach to hospitalizationsnrelated to musculoskeletal injuries as defined by the ICD-10 clas-ification of diseases. Because the PDHS is primarily focused onusculoskeletal injuries we would generally not expect to observe

n effect on other types of health conditions. Appendix shows thathe point estimates are all economically very small and predomi-antly insignificant. Generally, these results suggest that the effectseported in Table 3, which we attribute to the PDHS are not causedy a general health trend.

As an additional sensitivity check we estimate the effects ofhe PDHS using propensity score matching methods (PSM). Theim is to ensure that the distribution of outcomes and backgroundharacteristics are similar in pre-program years. Table 5 presentsSM estimates for the combined treatment group separately forach pre- and post-program year. Rows (1)–(3) show that match-ng effectively eliminates any treatment-control differences in there-program outcomes. The PSM estimates for post-program yearsre reported in rows (4)–(6). The general pattern of results is sim-

lar to that found in Fig. 3, thus confirming the robustness of our

ain findings.Finally, to partially gauge the potential impact of the sam-

le selection bias in the job absenteeism estimates, Table A.3 in

132 M.S. Pedersen, J.N. Arendt / Journal of Health Economics 37 (2014) 123–136

-0.4-0.3-0.2-0.1

00.10.20.30.4

2005 2006 2007 2008 2009 2010

A. Physician contacts (count)

-0.04-0.03-0.02-0.01

00.010.020.030.04

2005 2006 2007 2008 2009 2010

B. Medicati on use (ATC M) (in%)

-0.04-0.03-0.02-0.01

00.010.020.030.04

2005 2006 2007 2008 2009 2010

C. Hospitalizati ons (I CD-10 M) ( in%)

-0.04-0.03-0.02-0.01

00.010.020.030.04

2005 2006 2007 2008 2009 2010

D. Hospitalizations (ICD-10 S) (in%)

-0.04-0.03-0.02-0.01

00.010.020.030.04

2005 2006 2007 2008 2009 2010

E. Job absenteeism (>3w) (in%)

-0.04-0.03-0.02-0.01

00.010.020.030.04

2005 2006 2007 2008 2009 2010

F. Job absenteeis m (>7w) (in%)

-0.04

-0.02

0

0.02

0.04

0.06

0.08

0.1

2005 2006 2007 2008 2009 2010

G. Ln(total income)

-0.04-0.03-0.02-0.01

00.010.020.030.04

2005 2006 2007 2008 2009 2010

H. Prescribed physical therapy (in%)

-0.04-0.03-0.02-0.01

00.010.020.030.04

2005 2006 2007 2008 2009 2010

I. Any chiropractic care (in%)

F he proe ation.M conne

Awetrj

5

mcpowbtuab

aaisdsbeidoipwhether workers were permanently disabled in 2010. In this case,the municipality may refer the workers to wage-subsidized jobswith tasks adjusted to the reduced working capacity and with

ig. 3. Difference-in-differences estimates by year. Notes: The solid line indicates tffect using estimated robust standard errors adjusted for within-individual correl: Musculoskeletal system, ICD-10 M: Diseases of the musculoskeletal system and

ppendix presents results from difference-in-differences modelshere job absenteeism is imputed for those individuals who leave

mployment during the post-program period and where job absen-eeism is thus not observed. As shown in the table, the results areather insensitive to the inclusion of individuals with non-observedob absenteeism.

.5. Additional results on employment status

In this section, we present some additional results on employ-ent status in the year 2010 (the last period of our panel). By

onstruction, all sample members are employed in the entire pre-rogram period (2005–2007). Based on the small positive effectsn total income, we might expect that the PDHS would also affecthether workers are retained in their pre-program jobs either

ecause it improved their work capacity or because they value

he program. Because PDHS enrollment is contingent on contin-ed employment in a job that is covered by the relevant collectivegreement area, workers who place high value on the program maye motivated to remain in their pre-program jobs that provide u

gram effect, and the dashed lines represent a 95% confidence interval around the All estimates are relative to 2007, which is normalized to zero in the figures. ATCctive tissues, ICD-10 S: Injury.

ccess to the program. The value of the PDHS may vary widelycross workers, being particularly high, for example, for workersn poor initial health. To investigate this possibility, we constructeveral measures of the transition out of the pre-program job. Too so, we use the fact that our data identifies main employmenttatus at one particular point in the year. First, we construct ainary indicator for whether the workers remain full-time wagearners in the last year of our panel.12 We also construct a binaryndicator for whether workers had left the pre-program firm. Weefine as having “left” those who either took job at another firmr left the labor market (i.e., were not linked to a particular firmn 2010). Similarly, we look at whether workers had left the pre-rogram occupation (e.g., electrician) in 2010. Finally, we examine

12 The majority of individuals no longer wage earners in 2010 transitioned tonemployment and a few individuals transitioned to self-employment.

M.S. Pedersen, J.N. Arendt / Journal of Hea

Tab

le

5M

atch

ing

esti

mat

es

by

year

.

Dep

end

ent

vari

able

Phys

icia

n

con

tact

s(1

)M

edic

atio

n

use

(ATC

M)

(2)

Hos

pit

aliz

ed(I

CD

-10

M)

(3)

Hos

pit

aliz

ed(I

CD

-10

S)

(4)

Job

abse

nte

eism

(>3

wee

ks)

(5)

Job

abse

nte

eism

(>7

wee

ks)

(6)

Ln(t

otal

inco

me)

(7)

Pres

crib

ed

ph

ysic

alth

erap

y

(8)

An

y ch

irop

ract

icca

re

(9)

A. P

re-p

rogr

am

year

s1.

2005

−0.0

30

(0.0

49)

−0.0

00

(0.0

04)

−0.0

01

(0.0

02)

−0.0

01

(0.0

04)

0.00

0

(0.0

03)

−0.0

02

(0.0

02)

0.00

5

(0.0

04)

−0.0

01

(0.0

03)

−0.0

02

(0.0

04)

2.

2006

−0.0

24

(0.0

51)

−0.0

01

(0.0

05)

−0.0

00

(0.0

02)

0.00

1

(0.0

04)

0.00

0

(0.0

03)

0.00

1

(0.0

02)

0.00

3

(0.0

03)

−0.0

00

(0.0

03)

−0.0

03

(0.0

04)

3.

2007

−0.0

32

(0.0

52)

0.00

0

(0.0

05)

0.00

0

(0.0

02)

−0.0

01

(0.0

04)

0.00

1

(0.0

03)

0.00

1

(0.0

02)

0.00

1

(0.0

03)

−0.0

02

(0.0

03)

−0.0

03

(0.0

04)

B.

Post

-pro

gram

year

s4.

2008

0.02

7

(0.0

57)

−0.0

04

(0.0

05)

−0.0

04

(0.0

03)

−0.0

13

(0.0

04)**

*−0

.009

(0.0

03)**

*−0

.007

(0.0

02)**

*0.

030

(0.0

03)**

*−0

.009

(0.0

03)**

*−0

.011

(0.0

04)**

*

5.

2009

−0.0

30

(0.0

59)

−0.0

04

(0.0

05)

−0.0

04

(0.0

03)

−0.0

16

(0.0

04)**

*−0

.011

(0.0

03)**

*−0

.010

(0.0

02)**

*0.

077

(0.0

04)**

*−0

.015

(0.0

03)**

*0.

000

(0.0

04)

6.

2010

−0.0

20

(0.0

61)

−0.0

04

(0.0

05)

−0.0

04

(0.0

03)

−0.0

12

(0.0

04)**

*−0

.012

(0.0

03)**

*−0

.010

(0.0

03)**

*0.

094

(0.0

05)**

*−0

.017

(0.0

03)**

*0.

006

(0.0

04)*

Not

es: O

bser

vati

ons:

13,5

26

trea

ted

and

31,3

44

mat

ched

con

trol

s.

Prop

ensi

ty

scor

e

mat

chin

g

wit

h

a

0.00

5

rad

ius.

Boo

tstr

app

ed

stan

dar

d

erro

rs

wit

h

200

rep

lica

tion

s

are

give

n

in

par

enth

eses

. Var

iabl

es

use

d

in

the

esti

mat

ion

of

the

pro

pen

sity

scor

e

are

show

n

in

Ap

pen

dix

. Est

imat

ed

are

for

the

com

bin

ed

trea

tmen

t

grou

p

of

“plu

mbe

rs”

and

“ele

ctri

cian

s”. A

TC

M:

Mu

scu

losk

elet

al

syst

em, I

CD

-10

M:

Dis

ease

s

of

the

mu

scu

losk

elet

al

syst

em

and

con

nec

tive

tiss

ues

, IC

D-1

0

S:

Inju

ry.

*St

atis

tica

l sig

nifi

can

ce

at

the

10%

leve

l.**

Stat

isti

cal s

ign

ifica

nce

at

the

5%

leve

l.**

*St

atis

tica

l sig

nifi

can

ce

at

the

1%

leve

l.

rwticjo

ttmctccdaataaetesoafiptMmeppctpa

5

camtgemTesespsbp

us

lth Economics 37 (2014) 123–136 133

educed working hours.13 If workers with permanently reducedorking capacity are unable to remain in a wage-subsidized job,

hen the municipality may grant the workers a permanent disabil-ty pension that is financed entirely by the public authorities. Weonstruct a binary indicator for the transition to a wage-subsidizedob or entry into permanent disability pension using informationbtained from the DREAM database.

In Table 6, we present linear probability models of the effect ofhe PDHS on each employment outcome. Estimates are reported forhe full sample and for a subsample of workers with a pre-program

usculoskeletal injury and use of physical therapy or chiropracticare for whom the value of the PDHS and, thus, the cost of leavinghe pre-program jobs are likely to be particularly high. All modelsontrol for age, sex, marital status, native citizen status, number ofhildren in the household, skill and educational level, region of resi-ence, and firm size measured in 2007 before program adoption. Inddition, all models control for pre-program records of medicallyttended musculoskeletal injuries, physician contacts, physicalherapy and chiropractic care use, episodes of job absenteeism,nd total (logged) income. We find that enrollees are 3.7 percent-ge points more likely than the controls to remain full-time wagearners in 2010. Note that this might imply a small sample selec-ion bias in the estimates of the impact on job absenteeism, as thesestimates are conditional on being a wage earner. However, ashown in Section 5.4 this is not expected to have a large impactn the job absenteeism estimates. We also find that enrolleesre 4.0 percentage points less likely to have left the pre-programrm and 1.7 percentage points less likely to have left their pre-rogram occupation. Consistent with previous results, we find thathe effects are not universal across collective agreement areas.

oreover, we find that the effects are greater for workers whoight have valued the program; enrollees with a pre-program

pisode of a medically attended musculoskeletal injury are 4.0ercentage points more likely to remain a wage earner and 5.1ercentage points less likely to have left the pre-program firm. Byontrast, we do not have the statistical power or a sufficiently longime period to detect any significant associations with entry intoermanent disability either in the full sample or for workers with

pre-program musculoskeletal injury.

.6. Costs and benefits compared

We conclude our empirical results by crudely comparing theosts and estimated effects of the PDHS. As we noted earlier, thennual premium of the program was $5514 per enrollee, whichay ultimately be borne by the worker as part of the compensa-

ion package. At an assumed marginal tax rate of 47%, the annualovernmental income tax revenues foregone as a result of the taxxemption would amount to $26 per enrollee given that the pre-ium would remain the same if there had been no tax exemption.

o determine whether these costs are worthwhile for workers andmployers as well as for public budgets, we use the most con-ervative statistically significant difference-in-differences pointstimates for the combined treatment group. From the worker per-pective, the main benefit consists of a small positive effect on totalost-tax income. If we use the average annual total income in our

ample, and the assumed average tax rate of 47%, then the averageenefit amounts to $67,000 × 0.53 × 0.015 ≈ $530. This far sur-asses the average costs assuming that they are placed on workers.

13 Danish labor market policies include other types of wage subsidies in whichnemployed individuals are hired for a temporary period on ordinary terms. Theseubsidized jobs are not included in the outcome.14 1 US $ = 5.4 DKK.

134 M.S. Pedersen, J.N. Arendt / Journal of Health Economics 37 (2014) 123–136

Table 6Linear probability model estimates of PDHS enrollment effects on employment status in 2010.

Dependent variable:

Full-time wage- earner (1) Left pre-program firm (2) Left pre-program occupation (3) Permanent disability (4)

A. Full sample1. All 0.037 (0.003)*** −0.040 (0.005)*** −0.017 (0.003)*** −0.001 (0.001)2. “Plumbers” 0.034 (0.005)*** −0.020 (0.008)*** −0.020 (0.004)*** 0.001 (0.001)3. “Electricians” 0.038 (0.003)*** −0.051 (0.006)*** −0.015 (0.003)*** −0.001 (0.002)Mean of dep.var. 0.922 0.341 0.204 0.011No. of obs. 44,959

B. Pre-program musculoskeletal injury4. All 0.040 (0.006)*** −0.051 (0.009)*** −0.027 (0.005)*** −0.002 (0.003)5. “Plumbers” 0.028 (0.008)** −0.038 (0.013)*** −0.026 (0.006)*** 0.000 (0.004)6. “Electricians” 0.048 (0.007)*** −0.059 (0.011)*** −0.028 (0.006)*** −0.003 (0.003)Mean of dep.var. 0.912 0.356 0.216 0.016No. of obs. 14,275

C. Pre-program physical therapy or chiropractic care use7. All 0.041 (0.007)*** −0.026 (0.013)** −0.020 (0.006)*** −0.003 (0.003)8. “Plumbers” 0.022 (0.011)** 0.000 (0.014) −0.018 (0.009)** −0.001 (0.003)9. “Electricians” 0.053 (0.008)*** −0.032 (0.011)*** −0.020 (0.007)*** −0.005 (0.003)Mean of dep.var. 0.922 0.353 0.214 0.018No. of obs. 7,418

Notes: Robust standard errors are given in parentheses. All models control for age, sex, marital status, number of children in household, region of residence, educational attain-ment, skill level, ln(total income), pre-program physical therapy and chiropractic care use, pre-program musculoskeletal injury, job absenteeism, labor union membership,a

FreaitfolIeea0etrcae

6

mwihacattt

w

cfatintlw

Fumctvapp

A

aKZnwwPA

nd firm size in 2007.* Statistical significance at the 10% level.

** Statistical significance at the 5% level.*** Statistical significance at the 1% level.

rom the employer perspective, the main benefit consists ofeduced costs associated with job absenteeism. Given that employ-rs paid sick-leave benefits only for the first 3 weeks of absence,15

lower bound of the absenteeism cost per enrollee would be $740n weekly sick-leave benefits × 3 weeks × 0.005 ≈ $11 although therue cost might well be larger if the employer tops up the benefit orace transaction costs during worker absence. From the perspectivef the public budgets, the main benefits consist of reduced sick-eave payments and increased governmental income tax revenues.f we use only the first 8 weeks of absence, subtracting themployer-financed period, then the PDHS relieves the public budg-ts of at least $740 × (8 − 3) × 0.005 ≈ $19 per enrolled worker. Inddition, the positive effect on total income translates to $67,000 ×.47 × 0.015 ≈ $470 in higher annual government income tax rev-nues per enrollee. If we add to that a reduction in marginal excessax burden resulting from the lower need for financing health-elated job absenteeism as well as the small reduction in physicianontacts and physician prescribed physical therapy, the PDHS thenppears cost-effective for the public sector, firms, workers and soci-ty as a whole in the short and medium term.

. Conclusion

This paper presents empirical evidence of the short- andedium-term effects of the PensionDanmark Health Scheme,hich provides prevention and early management of work-related

njuries. The evidence presented here suggests that the programad modest effects (if any at all) on the prevention of medicallyttended injuries in the first 3 years after introduction in theonstruction sector. Interestingly, however, we find evidence of

reduction in episodes of health-related job absenteeism condi-

ional on employment and a small positive effect on total incomehat more than outweighs the program costs. A number of fac-ors lead us to believe that the estimated effects are most likely

15 As noted previously, the public sector covers sick-leave benefits beyond threeeeks of absence.

gHvUarwi

ausal. First, we are able to avoid potential self-selection issues byocusing on a program that was adopted at the level of collectivegreements and was mandatory for all covered workers. Second,he treatment and control groups generally showed similar trendsn the pre-program period. Third, the program generally showedo effects on outcomes that should not have been affected. Fourth,he effects of the program increased by firm size, possibly becausearge firms have more resources and social networks to motivate

orkers to engage with the program.Naturally, our empirical analysis contains some limitations.

irst, there might be longer-term effects than what the data alloweds to investigate in this study. In addition, we did not have infor-ation on other health-promoting efforts in the workplace that

ould have simultaneously affected the outcomes. However, forhese confounding factors to bias our results, they would need toary in the same timely pattern across the collective agreementreas as the program. At a minimum, the results presented in thisaper suggest that widespread and relatively low-cost PDHS-stylerograms warrant further study.

cknowledgments

For helpful comments, the authors would like to thank twononymous referees, Mickael Bech, Nabanita Datta Gupta, Astridiil, Kjeld Møller Pedersen, Paul Sharp, Andrea Weber, and Peterweifel, as well as seminar participants at the Department of Busi-ess and Economics, University of Southern Denmark, the 5thorkshop of the Danish Health Econometrics Network, the 3rdorkshop of the Centre for Research in Active Labour Market

olicy Effects, Aarhus University, and the 2012 DGPE workshop.ny remaining errors are the responsibility of the authors. Specialratitude is also owed to the staff at PensionDanmark and Falckealthcare for unrestricted funding for this project and for pro-iding us with data and patiently answering numerous questions.nfortunately, the administrative pension records are proprietary

nd may not be made available to other researchers. Administrativeegister data can be obtained by any researcher with an affiliationith a Danish research institution. The authors have no conflict of

nterest in this work.

of Health Economics 37 (2014) 123–136 135

A

TD

Niroi

TE

Table A.2 (Continued)

Independent variable All

Treatment Control %Norm. diff.

General upper secondary school (%) 0.8 0.8 0.3Vocational education (%) 84.6 85.2 1.5Short-cycle education (%) 6.1 5.7 2.1Medium-cycle education (%) 0.6 0.6 0.1Long-cycle education (%) 0.0 0.0 0.0Marital status (%) 74.1 74.3 0.5Children in household (%) 0.9 0.9 0.0Capital Region of Denmark (%) 29.2 29.9 1.7Region Zealand (%) 19.6 20.0 1.2Region of Southern Denmark (%) 21.5 20.8 1.7Central Denmark Region (%) 19.1 18.9 0.6North Denmark Region (%) 10.7 10.4 0.8Labor union member (%) 92.1 91.8 1.1Manager (%) 0.4 0.4 0.2Salaried (%) 1.9 1.9 0.3Skilled (%) 94.3 94.4 0.3Unskilled (%) 0.8 1.1 0.9Other wage earner (%) 2.5 2.2 1.2Firm size 1–10 employees (%) 17.6 17.6 0.1Firm size 10–30 employees (%) 31.7 32.5 1.8Firm size 30–100 employees (%) 23.8 23.8 0.0Firm size >100 employees (%) 26.9 26.1 1.9

No. of obs. 13,526 31,344

Notes: Propensity score matching with 0.005 radius. A total of 14 individuals wereoutside the area of common support. % Norm diff: Percentage normalized difference,which is calculated according to 100 · x̄i1−x̄i0√

V1(xi )+V0(xi )/2where x̄i1 denotes the treated

unit mean for variable i, and V0(xi) denotes the variance of the variable i within thecontrol group. The variables are measured in 2007 unless otherwise noted.

Table A.3Differences-in-differences estimates with imputed job absenteeism forunemployed.

Dependent variable All

Jobabsenteeism(>3 weeks)

Jobabsenteeism(>7 weeks)

No. ofobs.

A. No linear time trend(s)Base-case results

(restricted to employed)−0.007 (0.002)*** −0.008 (0.002)*** 252,240

Job-absenteeism imputedto zero for unemployed

−0.008 (0.002)*** −0.009 (0.002)*** 273,476

Job-absenteeism imputedto the sample mean forunemployed

−0.006 (0.002)*** −0.007 (0.002)*** 273,476

Job-absenteeism imputedto 2 times the samplemean for unemployed

−0.006 (0.002)*** −0.007 (0.002)*** 273,476

B. With linear time trend(s)Base-case results

(restricted to employed)−0.001 (0.004) −0.005 (0.003)* 252,240

Job-absenteeism imputedto zero for unemployed

−0.004 (0.004) −0.008 (0.003)*** 273,476

Job-absenteeism imputedto the sample mean forunemployed

−0.003 (0.004) −0.006 (0.003)** 273,476

Job-absenteeism imputedto 2 times the samplemean for unemployed

−0.003 (0.004) −0.006 (0.003)** 273,476

Notes: Robust standard errors adjusted for within-individual correlation are givenin parentheses. All models control for age, sex, marital status, native citizen status,number of children in the household, region of residence, educational attainment,skill level, labor union membership, and firm size in 2007. Models with linear timet

M.S. Pedersen, J.N. Arendt / Journal

ppendix.

able A.1ifference-in-differences estimates on hospitalizations defined by the ICD-10.

Dependent variable All

No linear trend With linear trend

Certain infectious diseases (A + B) 0.001 (0.006)* 0.001 (0.001)Neoplasms and diseases of the blood

and blood-forming organs (C + D)−0.002 (0.001)* −0.003 (0.001)**

Endocrine, nutritional and metabolicdiseases (E)

0.000 (0.001) 0.002 (0.001)

Mental and behavioral disorders (F) 0.000 (0.000) 0.001 (0.001)Diseases of the nervous system (G) −0.000 (0.001) −0.001 (0.001)Diseases of the eye and ear (H) −0.003 (0.001)** 0.002 (0.002)Diseases of the circulatory system (I) −0.002 (0.001) −0.002 (0.002)Diseases of the respiratory system (J) 0.000 (0.001) −0.000 (0.001)Diseases of the digestive system (K) −0.001 (0.001) −0.000 (0.002)Diseases of the skin (L) 0.000 (0.001) −0.001 (0.001)Diseases of the urogenital system (N) 0.002 (0.001)* 0.004 (0.002)**

Pregnancy and childbirth (O) −0.000 (0.000) −0.000 (0.000)Certain conditions originating in the

perinatal period (P)−0.000 (0.000) −0.000 (0.000)

Congenital malformations (Q) −0.000 (0.000) 0.000 (0.001)Symptoms, sings and abnormal clinical

findings not classified elsewhere (R)−0.003 (0.001)** −0.003 (0.002)

Poisoning (T) −0.003 (0.001)** −0.003 (0.002)External causes of morbidity and

mortality (V + X + Y)−0.000 (0.000) −0.000 (0.000)

Factors influencing health status andcontact with health services (Z)

−0.004 (0.002) −0.003 (0.004)

No. of obs. 273,476

otes: Robust standard errors adjusted for within individual correlation are givenn parentheses. All models control for marital status, the number of children, theegion of residence, whether an individual has completed primary or lower sec-ndary education, job type, labor union membership, and the number of employeesn the firm. All models also include a linear treatment group-specific time trend.

* Statistical significance at the 10% level.** Statistical significance at the 5% level.

*** Statistical significance at the 1% level.

able A.2mpirical means by treatment and control groups after matching.

Independent variable All

Treatment Control %Norm. diff.

Physician contacts (#), 2007 3.7 3.7 0.7Physician contacts (#), 2006 3.5 3.5 0.5Physician contacts (#), 2005 3.3 3.3 0.7Medication use (ATC M) (%), 2007 18.7 18.7 0.0Medication use (ATC M) (%), 2006 17.7 17.8 0.2Medication use (ATC M) (%), 2005 17.0 17.0 0.1Hospitalized (ICD-10 M) (%), 2007 4.5 4.5 0.0Hospitalized (ICD-10 M) (%), 2006 3.8 3.8 0.0Hospitalized (ICD-10 M) (%), 2005 3.7 3.9 0.8Hospitalized (ICD-10 S) (%), 2007 13.4 13.6 0.4Hospitalized (ICD-10 S) (%), 2006 14.0 14.0 0.2Hospitalized (ICD-10 S) (%), 2005 14.6 14.7 0.3Job absenteeism >3 weeks (%), 2007 5.9 5.8 0.3Job absenteeism >3 weeks (%),2006 5.4 5.2 0.4Job absenteeism >3 weeks (%), 2005 5.1 5.1 0.1Ln(total income), 2007 12.8 12.8 0.5Ln(total income), 2006 12.8 12.8 1.0Ln(total income), 2005 12.7 12.7 1.9Physical therapy (%), 2007 6.4 6.6 0.6Physical therapy (%), 2006 6.0 6.0 0.0Physical therapy (%), 2005 5.7 5.8 0.4Any chiropractic care (%), 2007 11.0 11.3 0.9Any chiropractic care (%), 2006 10.6 10.9 1.1Any chiropractic care (%), 2005 9.7 9.9 0.6

Age (years) 37.2 37.0 1.8Native (%) 97.4 97.2 1.6Male (%) 98.7 98.3 3.1Basic school 8–10th grade (%) 7.8 7.6 0.7

v

rends include a linear time trend variable interacted with the treatment group ariable.

* Statistical significance at the 10% level.** Statistical significance at the 5% level.

*** Statistical significance at the 1% level.

1 of Hea

R

A

A

A

B

B

B

B

B

C

C

D

D

F

F

G

G

G

G

H

H

H

H

K

K

L

L

L

L

M

M

P

P

R

R

S

SJournal of Human Resources 14, 145–170.

36 M.S. Pedersen, J.N. Arendt / Journal

eferences

akvik, A., Holmås, T.H., Kjerstad, E., 2003. A low-key social insurancereform—effects of multidisciplinary outpatient treatment for back pain patientsin Norway. Journal of Health Economics 22, 747–762.

ngrist, J., Pischke, J.S., 2008. Mostly Harmless Econometrics: An Empiricist’s Com-panions. Princeton University Press, Princeton, New Jersey.

uld, M.C., Emery, J.C.H., Gordon Daniel, V., McClintock, D., 2001. The efficacy ofconstruction site safety inspections. Journal of Labor Economics 19, 900–921.

aicker, K., Cutler, D., Song, Z., 2010. Workplace wellness programs can generatesavings. Health Affairs 29, 304–311.

ureau of Labor Statistics (BLS), 2013. OS NR 11/07/2013 News Release: WorkplaceInjuries and Illnesses—2012.

oden, L.I., Galizzi, M., 2003. Income losses of women and men injured at work.Journal of Human Resources 38, 722–757.

oden, L.I., Ruser, J.W., 2003. Workers’ compensation reforms, choice of medical careprovider, and reported workplace injuries. Review of Economics and Statistics85 (4), 923–929.

utler, R.J., Baldwin, M.L., Johnson, W.G., 2006. The effects of occupational injuriesafter return to work: work absences and losses of on-the-job productivity. Jour-nal of Risk and Insurance 73 (2), 309–334.

ase, A., Deaton, A.S., 2005. Broken down by work and sex: how our health declines.In: Analyses in the Economics of Aging. University of Chicago Press, Chicago.

awley, J., Price, J.A., 2013. A case study of a workplace wellness program that offersfinancial incentives for weight loss. Journal of Health Economics 32, 794–803.

anish Ministry of Taxation, 1995. Arbejdsgiverbetalte sundhedsudgifter. Ministryof Taxation, Betænkning nr. 1296/Juni 1995 (in Danish).

ehejia, R.H., Wahba, S., 2002. Propensity score-matching methods for nonexperi-mental causal studies. Review of Economics and Statistics 84, 151–161.

letcher, J.M., Sindelar, J.L., Yamaguchi, S., 2011. Cumulative effects of job charac-teristics on health. Health Economics 20, 553–570.

rölich, M., Heshmati, A., Lechner, M., 2004. A microeconometric evaluation of reha-bilitation of long-term sickness in Sweden. Journal of Applied Econometrics 19,375–396.

alizzi, M., Boden, L., 2003. The return to work of injured workers: new evi-dence from matched unemployment insurance and workers’ compensationdata. Labour Economics 10, 311–337.

upta, N.D., Lau, D., Pozzoli, D., 2012. The impact of education and occupation ontemporary and permanent work incapacity. IZA Discussion Paper Series, No.6963.

ustman, A.L., Steinmeier, T.L., 2005. Imperfect knowledge of social security andpensions. Industrial Relations: A Journal of Economy and Society 44, 373–397.

uzman, J., Esmail, R., Karjalainen, K., Malmivaara, A., Irvin, E., Bombardier, C., 2001.

Multidisciplinary rehabilitation for chronic low back pain: systematic review.British Medical Journal 322, 1511.

eckman, J.J., Hotz, V.J., 1989. Choosing among alternative nonexperimental meth-ods for estimating the impact of social programs: the case of manpower training.Journal of the American Statistical Association 84, 862–874.

T

Z

lth Economics 37 (2014) 123–136

øgelund, J., Holm, A., 2006. Case management interviews and the return to workof disabled employees. Journal of Health Economics 25 (3), 500–519.

øgelund, J., Holm, A., McIntosh, J., 2010. Does graded return-to-work improve sick-listed workers’ chance of returning to regular working hours? Journal of HealthEconomics 29 (1), 158–169.

uber, M., Lechner, M., Wunsch, C., 2013. The performance of estimators based onthe propensity score. Journal of Econometrics 175 (1), 1–21.

enkel, D., Supina, D., 1992. The determinants of worksite health promotion. Eco-nomics Letters 40, 345–351.

rueger, A., Rouse, C., 1998. The effect of workplace education on earnings, turnoverand job performance. Journal of Labor Economics 16, 61–94.

anoie, P., 1992. The impact of occupational safety and health regulation on the riskof workplace accidents: Quebec, 1983–1987. Journal of Human Resources 27,643–660.

aun, L., Thoursie, P.S., 2014. Does privatisation of vocational rehabilitation improvelabour market opportunities? Evidence from a field experiment in Sweden.Journal of Health Economics 34, 59–72.

eigh, J.P., 2011. Economic burden of occupational injury and illness in the UnitedStates. Milbank Quarterly 89, 728–772.

ipscomb, H.J., Dement, J.M., Silverstein, B., Cameron, W., Glazner, J.E., 2009. Whois paying the bills? Health care costs for musculoskeletal back disorders,Washington State Union carpenters 1989–2003. Journal of Occupational andEnvironmental Medicine 51, 1185–1192.

eyer, B.D., Viscusi, W.K., Durbin, D.L., 1995. Workers’ compensation and injuryduration: evidence from a natural experiment. American Economic Review 85(3), 322–340.

orefield, B., Ribar, D., Ruhm, C., 2012. Occupational status and health transitions.BE Journal of Economic Analysis & Policy 11, 1–29.

uhani, P.A., Sonderhof, K., 2010. The effects of a sick pay reform on absence and onhealth-related outcomes. Journal of Health Economics 29, 285–302.

unnett, L., Wegman, D.H., 2004. Work-related musculoskeletal disorders: theepidemiologic evidence and the debate. Journal of Electromyography and Kine-siology 14, 13–23.

oos, E., Bliddal, H., Christensen, R., Hartvigsen, J., Mølgaard, C., Søgaard, K., Zebis,M.K., 2013. Forebyggelse af skalder og sygdomme i muskler og led. Vidensrådfor forebyggelse, Available at: vidensraad.dk (in Danish).

osenbaum, P.R., Rubin, D.B., 1983. The central role of the propensity score in obser-vational studies for causal effects. Biometrika 70, 41–55.

chonstein, E., Kenny, D., Keating, J., Koes, B., Herbert, L., 2003. Physical conditioningprograms for workers with back and neck pain: a cochrane systematic review.Spine 28, E391–E395.

mith, R.S., 1979. The impact of OSHA inspections on manufacturing injury rates.

veito, T.H., Hysing, M., Eriksen, H.R., 2004. Low back pain interventions at theworkplace: a systematic literature review. Occupational Medicine 54 (1), 3–13.

iebarth, N.R., Karlsson, M., 2010. A natural experiment on sick pay cuts, sicknessabsence, and labor costs. Journal of Public Economics 94, 1108–1122.