Profiting from Presenteeism? Effects of an
Enforced Activation Policy on Firm Profits
Anna Godøy∗
December 11, 2014
Abstract
Activation requirements and graded benefits are a strategy for reducing social insurance costs in comprehensive welfare states. In Norway, a
policy of issuing graded rather than full time sickness absence certificates,
is viewed as a strategy not just to reduce direct costs of sick pay but also
to facilitate returns to work and reduce inflows to permanent disability.
This paper analyzes effects of partial sick leave on firm profits, on average and across different firm groups. An instrumental variable approach
exploiting variation in physicians’ certification practices is implemented
to identify causal effects of graded absence. On average, graded sickness
absence is found to increase firm profits at high skill firms; the effects are
strongest at the top of the productivity distribution. For low skill firms,
grading has no significant effects on profitability.
Keywords: absenteeism, productivity, activation, work loss
JEL Classification Numbers: I18, I38, J30, J48
1
Introduction
Activation policies have become a feature of different social insurance arrangements. In the case of sick pay, activation requirements in the form of graded
absence certificates have been found to reduce the time it takes to find work, as
well as reduce inflow to permanent disability benefits. There is limited evidence
on the effect these kinds of activation policies have on the firm side. This paper
investigates the impact of graded sickness absence on the profitability of firms.
The a priori effects of such activation policies on firm profits are ambiguous. On the one hand, absenteeism is generally perceived to be costly for firms,
so reducing absenteeism by having workers come in part time could have positive effects on profits. This is most clearly the case in regimes where the firm
∗ Institute for Social Research,
P. box 3233 Elisenberg, 0208 Oslo, Norway.
[email protected]. The research has received support from the Norwegian Research
Council (grant number 227117). Data made available by Statistics Norway have been essential.
1
continues to pay wages to absent workers on temporary disability. Moreover,
the cost of absence is typically thought to be higher than direct wage costs,
due to disruption of the firm production process. As a result, firms tend to be
concerned with absence rates even when firms do not cover sick pay for temporarily disabled workers, i.e. when sick pay is covered by the government or
in systems without sickness benefits. Issuing graded rather than part time sickness certificates could reduce the disruption of the production processes, which
would increase firm profits. In this way, activation policies reducing the degree
of absence also reduce the cost of absence, improving firm outcomes.
On the other hand, more use of graded sickness absence would mean that the
workers who actually show up to work on average have worse health, and thus
may be less productive, that is, there may be costs associated with increased
presenteeism. Full time absence may be preferable for firms if absent workers
are less productive per hour. Graded absence certificates are often implemented
by a proportional reduction in working hours - that is a worker who is assigned a
50% absence certificate will work 50% of their contracted hours. Put differently,
with graded absence certificates, the employer might end up paying full price
for workers who, due to a temporary disability, are less productive than usual.
Finally, fulfilling such activation requirements during sickness absence may have
direct implementation costs for firms, for instance some medical conditions may
require the employer to adjust the physical working environment.
The impact of graded sickness absence on firms’ outcomes should be of interest to policymakers for a number of reasons. The policy of issuing graded rather
than full time absence certificates has been found to lead to shorter absences
and higher re-employment probabilities (Høgelund et al. 2010, Viikari-Juntura
et al. 2012, Markussen et al. 2012) reducing the cost of sickness absence to the
public. To get a more complete view of the economic consequences of grading,
we should also look at costs incurred by the employers. By doing this, we can
find out whether grading reduces the costs of health related absences to all involved parties, as opposed to simply shifting the costs of sickness absence from
the taxpayers to the individual firms.
If grading impacts firm profitability, the incentives of firms to manage absence might also change. In Norway, as in several European countries, sick
pay for long term absence spells is covered by the government. In this kind of
institutional environment, economic theory suggests we should expect a moral
hazard problem. As firms are insured against the direct costs of long term
sickness absence, they will exert too little effort in keeping absence rates low
(OECD 2010). Empirical evidence suggests this is indeed the case: Fevang et al.
(2014), using data from a Norwegian reform removing firms’ copay for pregnant
women’s sickness absences, found significant increases in absence rates. If grading sickness certificates increases the costs of absence to firms, this could be
one channel explaining the beneficial effects of grading on patients’ labor market outcomes. That is, as graded benefits become more common, long term
absences become more costly for firms, giving firms stronger incentives to reduce absence rates. Conversely, if graded absences are less costly, moral hazard
problems may become more important.
2
At the same time, if grading has an impact on firm profits, this may also
affect firms’ hiring strategies. The willingness of firms to hire workers with
poor health status likely depends on the perceived cost of future absence spells.
If partial sick leave increases the cost of sickness absence relative to full time
absence, this could make firms more reluctant to hire workers with a history of
long sickness absences.
As a first step in investigating this issue, we could compare profits of firms
with mostly full time sickness absence to firms with higher shares of partial
sickness absence. However, this is problematic for several reason. First, absence
spells are more likely to be full time if they are expected to last a short time.
For this reason, firms with higher shares of partial sickness absence will also
typically have higher total absence rates.
Even controlling for firm total absence rates and diagnosis groups, this exercise risks giving biased estimates if some types of firms have an easier time
accomodating part-time sickness absence than others. The decision of whether
a sickness absence spell should be full time or part time is typically made following meetings between the patient, the certifying physician and the employer;
the preferences of the employer and the nature of work are both likely to influence the grading decision. For instance, some highly productive firms may
have a higher requirement for how healthy a worker needs to be in order to be
productive. Controlling for worker characteristics, these firms would have low
rates of graded absence (as the threshold for returning to work would be high)
and high profitability, making it seem as though there is a negative relationship
between grading and profits even in the absence of any causal effect.
To identify the effect, I will exploit variation the practice style of physicians.
There is significant variation in the practice style of physicians when it comes
to issuing graded rather than full time absence certificates. Linking employees
to physicians, we can estimate effects of the average practice style of the workers’ doctors on the profits of the firm. Some firms have workers whose doctors
regularly prescribe part-time rather than full time sickness certificates. Other
firms’ workers may have doctors who - on average - prefer to issue full time absence certificates. Comparing profits across firms with different types of doctors
should help uncover the causal effects of activation on firm profitability.
The identification strategy used in the present paper was first introduced by
Markussen et al. (2012), who used variation in physicians’ certification practices
to identify effects of activation requirements during long term certified sickness
spells on individual labor market outcomes. The authors find that receiving
a partial sickness absence, rather than being absent full time, leads to shorter
absence spells and higher employment rates. In particular, the positive effects of
grading on the probability of returning to employment suggests that activation
grading might have a positive effect on firm profits, as the firm will not have to
incur costs of hiring and training new workers.
The present paper is related to a literature on the cost of sickness absence
to firms. If absence is costly to the firm above and beyond direct wage costs,
we might expect part-time absences to reduce some of these costs. Pauly et al.
(2002) present a theoretical model of the costs of employee absence. Absence
3
is found to be more costly when there is team production and where substitute
workers are difficult to find. In these cases, the firm’s cost of absence will be
higher than the wage costs. Nicholson et al. (2006) use survey data on perceived
costs of absences and job characteristics to empirically evaluate the implications
of this model. The authors find that absence costs typically exceed wage costs,
with an average figure of 28%. When disaggregating results by profession, excess costs were found to be particularly high for high skill occupations (e.g.
engineers), and lower skill jobs working closely with high skill professionals (e.g.
paralegals). For low skill jobs (fast food cooks, waiters etc,) there was little excess cost of absence above the wage rate. From their findings, we should expect
positive effects of grading, in particular for high skill firms.
There is also a large related literature concerned with measuring the impact
of various health conditions on firm performance. These studies typically focus
on the impact of specific medical conditions, for example depression (Stewart
et al. 2003, Simon et al. 2001), back pain (van Tulder et al. 1995). Collins et al.
(2005) use survey data from a single US corporation to assess the cost of chronic
health conditions to the firm from treatment costs and output loss from absenteeism and on-the-job productivity loss (”presenteeism”). The authors conclude
that the cost of presenteeism are large and higher than the costs associated with
absences. These studies are focused on a somewhat different question than the
present paper - attempting to estimate the full impact of health conditions, taking into account treatment costs - whereas the present paper seeks to identify
effects on the firm of absence patterns for a given (initial) health status.
The contribution of this paper is twofold. First, it contributes to the literature on costs of sickness absence to firms, by analyzing detailed register data
on the full population of firms. Second, it contributes to the literature on activation strategies in social insurance arrangements by explicitly discussing the
role of the firm.
The rest of the paper is organized as follows. Section 2 gives a brief overview
of the relevant institutional context and presents the data used in the analysis.
The physician grading propensity instrument is described in section 3. Section
4 presents the empirical models and the estimated effects, and section 5 presents
some concluding remarks.
2
Institutions and data
In Norway, persons who are unable to work due to sickness can receive sickness
benefits at 100% replacement ratio for sickness spells lasting up to 12 months.
Workers who are absent up to 12 months benefit from additional job protection,
meaning they cannot be fired for reasons related to their absence. While very
short absences - typically up to three days - can be self certified, absences lasting
longer than this need to be certified by a doctor. The first 16 days of each spell
sickness benefits are paid by the employer, after that benefits are paid by the
government. In other words, firms incur the largest direct costs of sickness
absence for relatively short sickness absence, while the direct costs of long term
4
absences will be considerably lower. After 12 months, the worker is no longer
entitled to sick pay, but is eligible for other benefits with significantly lower
replacement ratios (typically around 66%). Job protection is also reduced after
12 months absence.
Employers are required by law to make necessary accommodations to facilitate return to work. For longer absence spells, the responsibilities of the
firm become more involved: Within four weeks, the employer should provide a
”follow-up”-plan. At 8 weeks, the certifying physician and the Labour and Welfare Administration (NAV, the agency responsible for paying sickness benefits)
meet to formally discuss possibilities of returning to work.
Absence certificates will state the degree of disability; the minimum absence
grade is 20%. A worker who is issued a graded absence certificate -that is, less
than 100% disability - can comply in two ways: The first, and most common way,
is by working the number of contracted hours multiplied by the assigned grade,
at full capacity. For example, a full time worker who is issued a 40% absence
certificate could fulfill this requirement by working two full days a week. The
second way a graded absence can be implemented is by working a larger number
of hours at a reduced work capacity, e.g. a full time worker can be issued a 40%
absence certificate, be present at work four full days a week and work at 50%
capacity.
In this paper, data on physician certified absences is merged to data on firm
performance to shed some light on how these activation -requirements affect
firm profits. The main sample is a firm-year dataset of annual profits, firm
characteristics and sickness absence rates. The starting point of the sample is
data from 2004 to 2010 on all private sector limited liability companies where
accounting data is available. Using this accounting data, the return on assets
(ROA) is defined as operating profits divided by total assets. This data is linked
to information on the firm’s location and industry classification.
The sample is then linked to demographic and educational data on the workers. This data includes information about gender, age, education and immigrant
status. Finally, data on doctor-certified sickness absence is added to the sample.
Absences are classified as full time if the disability rate is 100%, and part time
if it is less than 100%.
Firm-years with missing data are excluded from the sample. The final sample
contains 167,107 firm-year observations. Table 1 shows some summary statistics
of this firm level sample. A large majority, 91% of firms experience some sickness absence each year, with the average sickness absence rate being at 5.0%.
Meanwhile, about 50% of firms experience any part-time sickness absence.
These numbers mask substantial variation in absence patterns between firms.
Specifically, there appear to be differences in both the level of absence rates
between high and low skill firms. Figure 1 shows the evolution of the average
rate of sickness absence and the average share of absence days that are graded
during this period. In this figure, firms are classified according to the education
background of their employees. Firms are classified as ”high education” if more
than 33% of employees have at least a college degree, and low education if less
than 10% of workers have a college degree.
5
Table 1: Summary statistics, main sample
Return on assets
Employment
Fraction female
Fraction immigrant
Avg employee age
Share college degree
Sickness absence (all)
Sickness absence (partial)
Any sick absence
Any partial sickness absence
Observations
mean
0.0715
20.30
0.377
0.0778
37.77
0.221
0.0498
0.0133
0.908
0.498
167107
sd
0.142
67.16
0.322
0.151
7.447
0.242
0.0515
0.0256
0.290
0.500
Note: Figure shows summary statistics of the panel of firms (2004-2010).
Figure 1: Absence rates, 2004-2010
Share of absence days graded
.15
.035
.04
.2
.045
.05
.25
.055
.06
.3
Average absence rates
2004
2006
2008
2010
2004
Year
2006
2008
2010
Year
Low education
All
High education
Note: Figure shows certified absence rates, calculated as the fraction of worker-years with any
sickness absence, not adjusted for grading. The share of absence graded is calculated as the
graded absence rate relative to the total average absence rate. Low education firms are firms
where 10% or less have a college degree; high education firms are firms where at least 33% of
workers have a college degree.
6
Figure 2: Grading and absence rates
0
.1
Share graded
.2
.3
.4
.5
Avg. share of absence graded and avg. absence rate
0
.05
.1
.15
Absence rate
.2
.25
Note: Figure shows a binned scatter plot of the share of absence that is graded and the average
absence rate in firms.
The left panel shows the evolution in average absence rates. High skill firms
have on average lower rates of sickness absence than low skill firms; the difference
is around 2 percentage points and appears to be stable over time. There has
been an upward trend in the use of graded absence certificates. In 2004, 17.7% of
the average firm’s absence days were associated with spells with graded absence
certificates; in 2010, that fraction had grown to 22.9%. Looking at the panel
on the right, graded absence certificates are used more often in high skill firms
than in low skill firms.
Initial costs associated with accommodating the needs of employees returning to work on a graded absence certificate might make doctors less likely to
issue graded absence certificates Figure 2 shows the relationship between the
average annual absence rate (not adjusted for absence grade) and the fraction
of absence days that are graded. There appears to be a positive relationship
between absence rates and the fraction of the absence that is graded, which
flattens out at high absence rates.
Finally, to motivate the formal analysis to follow, figure 3 plots the relationship between the share of absence graded and firm profits. Figure 3 plots return
on assets (ROA) and the share of long-term absence - absence excluding the first
16 days of each spell - that is graded. As discussed above, short term absence
has a higher direct cost to firms, and short lasting spells are also less likely to
be graded. If short term spells were included when computing the average share
of graded spells, we would capture in part the relationship between short term
absence and profits, rather than the relationship between profits and grading.
Figure 3 appears to indicate a positive relationship between grading and
profits. On average, firms where a larger share of long term absence is graded
have higher return on assets, compared to firms where a higher share of absence
is full time. Of course, this relationship could reflect many things and should
7
Figure 3: Return on assets and share graded
.065
.07
ROA
.075
.08
ROA and avg. share of absence graded
0
.2
.4
.6
Share graded
.8
1
ROA and avg. share of absence graded
>16 days only
.065
.065
.07
ROA
ROA
.07
.075
.075
.08
Up to 16 days
0
.2
.4
.6
Share graded
.8
1
0
.2
.4
.6
Share graded
.8
1
Note: Figure shows binned scatter plots of firms’ return on assets (ROA) and the share of
absence that is graded.
8
not be given a causal interpretation. Some of this likely reflects the way grading
is used more often in high skill firms (figure 1), which may be more productive.
Any analysis of this relationship should then at a minimum control for employees’ observable characteristics. Moreover, a higher share of graded absence
could reflect favorable firm characteristics such as better health of the workers,
better quality of management, less workplace conflicts and so on, that are not
observed in the available register data. One way to investigate this further will
be through an instrumental variable approach. The next section presents the
instrumental variable strategy.
3
Physician instrument
As discussed in the previous sections, the decision to issue a graded rather
than a full time absence certificate is typically not random, but correlated with
observed and unobserved characteristics of patients and their work situations.
Data presented in the previous section showed that the use of graded sickness
absence varies across firms in a non-random way: High skill firms have a higher
relative incidence of partial sickness absence than low skill firms. To identify the
causal effect of graded absence on firm profits, I will implement an instrumental
variable strategy. The instrumental variable strategy used follows the work of
Markussen et al. (2012), where the authors use variation in the grading practices
of primary physicians to identify the impact of grading on individual absence
durations and employment outcomes.
The basic premise of the identification strategy is that doctors would differ in
their intrinsic grading propensity, i.e. how likely they are to issue a graded rather
than a full time absence certificate, even when faced with identical patients.
The grading decision is largely left up to the individual physician’s discretion,
as there are few formal guidelines for what diagnoses should be graded. If
doctors have different views on the usefulness of graded absence, or differ in
their willingness to negotiate with less cooperative patients (or their employers),
this could potentially give rise to exogenous variation in the likelihood that an
individual absence is graded. A major challenge to this identification strategy is
that the intrinsic grading propensity is unobserved, but will have to be estimated
using available data on actual grading decisions.
To estimate the grading propensity, I look at all absence spells lasting 8
weeks or longer, observing whether or not there is some incidence of graded
sickness absence in the first 8 weeks of the spell. To avoid confounding leniency
and grading propensity, a cutoff is imposed where only spells lasting longer
than the minimum duration are included in the estimation sample, ignoring
grading decisions that happen after this minimum duration. As discussed previously, doctors are generally more likely to issue full time absence certificates
when absences are expected to be of short duration. A lenient doctor, who is
willing to extend absences for longer, will then conceivably be observed with a
higher share of absence spells graded than a stricter doctor, even if the two have
identical intrinsic grading propensities. Setting the cutoff at exactly 8 weeks is
9
rather arbitrary, but allows for straightforward comparison with the findings of
Markussen et al. (2012).
It should be noted that this procedure does not rule out correlation between
leniency and grading propensity, rather, it merely avoids confounding the two
when they are uncorrelated. A positive correlation is problematic as this could
imply that doctors with high grading propensity on average have healthier patients at 8 weeks absence duration. Markussen et al. (2012) find no evidence
that the two are positively correlated, rather, the correlation is negative (but
small).
There is substantial variation in the observed grading practices of physicians, however this does not necessarily reflect differences in grading propensity.
Rather, this can be the result of differences in the characteristics of the patients
and their employment situations. In the empirical models, early grading will
be modeled as a flexible function of patients’ observable characteristics in order to remove these composition effects. The remaining variation in grading
probability will be interpreted as the doctors’ intrinsic grading propensity.
In order to estimate the doctor effects, an auxiliary dataset is constructed
consisting of long term absence spells merged with doctor identifiers and demographic and employment characteristics of the patients. The starting point of
the sample is all long term absence spells - defined as spells lasting at least 56
days - starting between 2004 and 2010. These absence data include information
on diagnosis and whether there was any grading prior to 8 weeks. Absence
spells are linked via unique person identifiers to the physician at the start of the
spell. Finally, data on demographics (age, gender, marital status, immigration
background, neighborhood), education (level and field) and job characteristics
(occupation and industry) are attached to the sample. The sample only includes
employed persons aged between 20 and 65 at the start of the absence spell.
Table 2 shows some descriptive statistics of the auxiliary sample. The sample
contains a total of 1,394,676 long term absence spells, of which 27% are graded
at some time during the first 56 days. Comparing columns 2 and 3 reveals
some differences in the characteristics of patients whose spell is graded early
and patients who are absent full time throughout the first 8 weeks. Patients
experiencing early grading are more likely to be female, less likely to be immigrants, have somewhat higher earnings, and are more likely to have a college
degree. The total duration of the absence spells are also shorter for persons
with some early grading.
Formally, the early grading decision is modeled as a simple linear probability
model:
partj = xj β 0 + εj
(1)
Where the dependent variable partj is an indicator variable part is constructed equal to 1 if the spell is graded within the first 8 weeks of the spell,
and 0 otherwise. xj is a number of observable characteristics that can be thought
to influence the grading decision, through health status or work situation. xj
contains age (dummies for five-year bins), education (78 dummies representing
10
Table 2: Summary statistics, auxiliary sample
Age
Female
Immigrant
Earnings (1000 NOK)
High school
College
Psychological
Musculoskeletal
Duration
Observations
(1)
All
mean
42.84
0.633
0.0769
338.6
0.443
0.247
0.198
0.379
178.8
1394676
(2)
Early graded
mean
42.11
0.752
0.0655
355.5
0.401
0.336
0.193
0.372
158.9
377337
(3)
No early graded
mean
43.12
0.589
0.0812
332.4
0.459
0.215
0.200
0.382
186.2
1017339
Note: Figure shows descriptive statistics of the auxiliary sample of long term absence spells
(duration >= 56 days). Column 2 represents spells where there was some grading during the
first 56 days of the spell, while column 3 represents spells where all first 56 days of the spell
were certified full time absences.
level and field of study), immigration background, industry and occupation, log
earnings, dummies for calendar month of entry and diagnosis group. xj also includes a set of 13,733 neighborhood indicators. As patients are likely to choose
doctors whose offices are close to where they live, differences in outcomes across
doctors could reflect socioeconomic sorting on unobservables if there is spatial
sorting within cities. Using residence data on a detailed level allows us to control
for spatial sorting in a more convincing way.
Equation (2) is estimated by OLS and the residual ε̂j is calculated for each
spell. ε̂j captures both the impact of unobserved factors related to spell j (severity of illness, precise nature of work), and the practice style of the physician.
These residuals are used to construct physician practice indicators for all patients registered with a primary care doctor, including persons with no sickness
absence spells. For each individual l registered with physician k, the physician
practice indicator is based on the average residual ε̂ of the doctor’s others patients. Letting K denote the set of long term spells certified by of doctor k,
L the set of long term spells belonging to patient l, the average residual for
individual l at doctor k is defined as
ε̂¯kl =
1
Nk − Nlk
X
ε̂j
j∈K,j ∈L
/
Nk is the total number of long term spells certified by doctor k and Nlk is
the number of spells certified by doctor k to patient k. For an individual with
no long term sickness absence, this is simply equal to the unadjusted mean of
11
0
.1
.2
.3
.4
Figure 4: Grading indicators
1
2
3
4
5
6
7
8
9 10 11 12 13 14 15 16 17 18 19 20
Doctors
Firms
Note: Figure plots the distribution of estimated firm-level physician grading indicators across
doctors (dark gray) or firm-years (light gray). The bars show binned averages, with each bar
representing 5% of doctors/firm-years.
the residuals of doctor k’s long term absence spells1 .
To facilitate interpretation, these residuals are added to the average predicted share of graded absences, given by x̄β̂ 0 , where x̄ is the average values of
observables x in the estimation sample, defining the physician grading indicator
of person l with physician k, ψkl :
ψkl = x̄β̂ 0 + ε̂¯kl
(2)
Finally, the individual workers grading instruments are aggregated to form
a firm level instrument ψ̂it . The physicians’ grading practice is less likely to
influence the absence behavior - and possibly firm profits - for workers who do
not have any sickness absence. On the other hand, current absence rates could
be influenced by previous years’ grading decisions. In the main specification, is
defined as the average values of all workers in firm i year t who entered some
certified sickness absence of any duration in the four years from t − 3 to t.
This aggregation is illustrated in figure 4, which plots the distribution of
physician grading indicators across individual physicians and across firm-years
in the sample. The dark gray bars show grouped averages of grading indicators
among doctors (ε̂¯k ), while the light gray bars show similarly constructed grouped
averages of (aggregated) grading indicators across firm-years (ψ̂it ). The columns
1 Excluding individual l’s own residuals from the calculation of physician averageε̂
¯kl in this
way removes a source of bias that would otherwise occur through the direct effect of l’s own
grading outcome on the profits of their employer. A similar bias could occur in a more indirect
way if co-workers tend to cluster around the same physician. In a robustness check to address
this concern, alternative physician averages are constructed where the residuals of all patients
who are working at individual l’s company are excluded from the calculation of ε̂¯kl ; estimated
effects remain virtually unchanged.
12
labeled 1 represent the average grading indicators in the group containing the
lowest 5% of doctors/firm-years, the columns labeled 2 represent the average
grading indicators in the group containing the next 5% of doctors/firm-years
etc.
As the firm-level grading indicators are essentially weighted averages of the
doctor level indicators, we would expect the firm level distribution to be less
dispersed. As indicated by figure 4, this is indeed the case, in particular at
the extremes of the distributions. The interquartile range of the firm level
instrument is 2.41 percentage points - a degree of dispersion that, while smaller
than the interquartile range of the distribution across doctors (6.33 percentage
points), is not insignificant. Meanwhile, a few firms have associated physicians
with predicted grading probabilities that are very low (including some below
zero) or very high. To avoid these few observations driving the results, firms
with the lowest and highest two percent of this distribution are excluded from
the estimation samples.
An objection to the use of this instrument is that patients are free to select
their doctors. If patients ”shop around” for doctors to find one that matches
their preference for full time versus graded absence, IV estimates may be contaminated by sorting effects. As noted by Markussen et al. (2012), sorting is
not necessarily problematic; as long as patients sort themselves to doctors according to observable characteristics that are included in the dataset, sorting
can be controlled for in the empirical models. However, if patients sort according to unobserved characteristics, such as ambition, health or willingness
to work, estimates will be biased. In the empirical section, this is addressed
through implementing robustness tests. In one robustness test, the grading
instrument is calculated using only workers who did not enter any certified sickness absence during the four years from t − 3 to t. If their physicians’ grading
practice are found to influence firm profits, this can be interpreted as evidence
of uncontrolled-for sorting.
As discussed above, grading propensity may be correlated with other aspects
of physician practice style, i.e. leniency. Failing to properly account for this
could lead to omitted variable bias of the IV estimates. For instance, if doctors
with a high grading propensity on average are less lenient, firms with high values
of the grading instrument ψ̂it would be expected to have lower absence rates
on average, all other things being equal, as their employees are less likely to
be granted absence certificates in the first place. This could in turn bias the
correlation between firm profitability and ψ̂it upward.
To address this, I follow Markussen et al. (2012) and estimate physician
indicators of physician leniency, estimating versions of (2) for two additional
outcomes: annual probabilities of receiving one or more sickness absence certificates (of any duration), and receiving one or more long term absence certificates
(absent for 8 weeks or longer). These equations are estimated using a personyear panel of all physicians’ registered patients aged 20-65 who are employed.
That is, the sample includes patient-years without sickness absence. Physicians are matched to their patients at the beginning of each calendar year. The
models include all the controls of equation (2), with the exception of diagnosis
13
categories (as these are not defined for persons with zero absence).
Having estimated these regressions, the predicted residuals are aggregated
by doctor and firm as outlined above. As the sample size is much larger (as it
includes all employed patients, not only on long term absence), physician averages are not corrected for own absences. However, physicians with less than 200
registered employed patients are excluded when computing firm averages. The
firm indicators are calculated using the average of all employees, without conditioning on absence history. These estimated leniency indicators are included
as controls in extended models.
4
Empirical models
To get a first impression of the data, I estimate a simple model of firm profits
yit = xit β x + θt + sickit β s + partit β p + εit
(3)
where yit is the profits of firm i year t (measured by return on assets). xit
is a vector of firm observable characteristics: the share of female workers, share
of immigrant workers, average age of workers and number of employees. θt are
year dummies.
The main variables of interest, sickit and partit represent the average incidence of sickness absence (total) and partial sickness absence, not adjusted for
certification degree. That is, β p captures the differential effect of partial sickness absence on profits: a positive estimate of β p indicates an that part-time
absence is less costly for the firm compared to full time absence. The upper
panel of table 3 contains estimated parameters from estimation of (3) on the
full sample.
A problem with estimating (3) by simple OLS is that both variables of interest sickit and partit are likely to be endogenous. If unobserved characteristics of
the firm influence both firm profits and sickness absence, as well as the proportion of sickness absence that is full time, the estimates will suffer from omitted
variable bias. To the extent that these unobserved characteristics are constant
over time, the problem can be mitigated by including firm fixed effects. Estimates from this specification are shown in column 2 of table 3. These models
find that sickness absence in general is costly for firms. However, for graded
absences, there is an additional positive effect. In the fixed effects specification,
this additional effect is slightly larger than the main effect of absence, indicating
that absence has no negative effects on firm profits as long as it is graded.
This interpretation is problematic however, as the specification does not take
into account the duration of absences. The lower panel of table 3 show estimates
from an alternative specification where I allow for differential effects of absence
that exceeds 16 days. For absences of this length, sick pay is covered by the
government and not by the firm, so the direct expenses will be smaller. In this
specification, long term absence is found to be less costly. The specification
without fixed effects finds significant differences in the costs of graded absence.
For short term absence, graded absence appears to be more expensive than full
14
Table 3: Effects of sickness absence rates
(a) Absence rates
Frac sick
Frac partial
Observations
(1)
OLS
-0.199∗∗∗
(-18.79)
(2)
Firm FE
-0.0881∗∗∗
(-8.77)
0.118∗∗∗
(5.91)
160244
0.0996∗∗∗
(5.28)
160244
t statistics in parentheses
∗
p < 0.10,
∗∗
p < 0.05,
∗∗∗
p < 0.01
(b) Long vs short term absence
(1)
OLS
-0.594∗∗∗
(-13.53)
(2)
Firm FE
-0.352∗∗∗
(-8.61)
Frac long time sick
0.463∗∗∗
(9.49)
0.303∗∗∗
(6.69)
Frac partial
-0.0999
(-0.77)
0.0643
(0.52)
Frac long time sick, partial
0.198
(1.42)
160244
0.0164
(0.13)
160244
Frac sick
Observations
t statistics in parentheses
∗
p < 0.10,
∗∗
p < 0.05,
∗∗∗
p < 0.01
Note: Models include controls for education background of the employees (share of employees
in each of 7 education levels), share of women, share of immigrants, average age, number of
employees, industry (68 categories), municipality and calendar year. Long term absence covers
all absence from day 17 of each spell, corresponding to the start of government refunded sick
pay. Standard error clustered at firm level.
15
time absence. For long term absences, the opposite appears to be true: the
additional effect of graded absence is positive. Both effects however are barely
significant. Moreover, when including firm fixed effects, estimated additional
effects of graded absence become smaller and statistically insignificant.
Including fixed effects does not account for unobserved heterogeneity that
changes over time. To account for this, I use an instrumental variable approach
exploiting residual variation in physicians’ grading propensity. The construction
of the physician practice style instrument is discussed in the previous section.
Consider the following model of firm profits:
yit = xit β + θt + sit γ + εit
(4)
where, as before, yit is return on assets and xit is a vector of firm observable
characteristics. sit is our main variable of interest and represents the share of
long term (> 16 days) sickness absence that was graded (any grade < 100%).
Table 4 shows selected estimates from model (4). The first panel contains
results from estimation of the model on the full sample of firms. From the OLS
estimates, the estimated effects of sit is positive and statistically significant.
Quantitatively, the effect is rather small. A one percentage increase in the share
of long term absence that is graded is associated with an increase of 0.0075
percent relative to the average ROA in the estimation sample (0.0696). The IV
estimates are presented in columns 2 and 3. The physician grading indicator
appears to be a strong instrument for the share of graded absence. Column
3 shows the estimated first stage: a one percentage point increase in the firm
average physician grading propensity increases the share of firm absences that
is graded by 0.617 percentage points. The instrumental variable approach does
not find any significant effects of the graded absence share on firm profitability.
The IV estimates are negative but not statistically significant.
The next set of models investigates whether the effect of graded absences
depends on the nature of work. As discussed in the previous sections, the use
of graded absence is higher in high education firms than in low education firms.
We would expect full time absence to be comparatively more costly when the
absent worker is difficult to replace, either by having the remaining workers
take over some of the responsibilities or by hiring a replacement worker on a
temporary contract. If work is more specialized at knowledge intensive firms, we
would expect the effect of activation to be larger for firms where more employees
have higher education degrees. The ”low education” category covers firms where
less than 10% of workers have a university degree, while the ”high education”
category are firms where at least 1/3 of workers have a university degree.
The estimated first stage is largely similar across the firm types, and the
instrument is strong with F of excluded instrument above 10 for both firm
categories. For low education firms, the IV estimates are slightly negative and
not significant. For high education firms, the IV estimates are positive and
significant. The IV estimates are also much larger in magnitude than the OLS
estimates: A one percent increase in the share of absence that is graded increases
return on assets by an average of 3.8 percent.
16
Table 4: IV estimates
(a) All firms
(1)
OLS
0.00523∗∗∗
(4.00)
Share partial (long)
(2)
IV
0.0429
(0.90)
Grading propensity
Observations
Kleibergen-Paap F-stat
118804
118804
85.59
(3)
First stage
0.617∗∗∗
(9.25)
118804
t statistics in parentheses
∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01
(b) Low education
(1)
OLS
0.00539∗∗∗
(2.64)
Share partial (long)
(2)
IV
-0.0173
(-0.28)
Grading propensity
Observations
Kleibergen-Paap F-stat
48851
48851
47.25
(3)
First stage
0.703∗∗∗
(6.87)
48851
t statistics in parentheses
∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01
(c) High education
Share partial (long)
(1)
OLS
0.00624∗∗
(2.41)
(2)
IV
0.303∗∗
(2.44)
Grading propensity
Observations
Kleibergen-Paap F-stat
27745
27745
17.95
(3)
First stage
0.582∗∗∗
(4.24)
27745
t statistics in parentheses
∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01
Note: Models include controls for education background of the employees (share of employees
in each of 7 education levels), share of women, share of immigrants, average age, number of
employees, industry (68 categories), municipality and calendar year. Low education firms are
firms where 10% or less have a college degree; high education firms are firms where at least
33% of workers have a college degree. Standard error clustered at firm level.
17
Grading appears to matter primarily for high skill firms. The models above
were estimated separately for small firms, at or below the sample median of 10
employees, and larger firms, with 11 or more employees. The IV estimates (not
reported) indicate no significant differential effects by firm size.
The IV estimate of the effect of graded absence on the profitability of high
skill firms seems large compared to the OLS estimates. There may be several
reasons for this: The first reason relates to some implicit sample selection issues: For firms with zero long term sickness absence, the dependent variable of
interest, share of absence graded, is not defined. As a result, these firms are
excluded from the regressions reported in table 4.
Second, the estimates above should be interpreted as local average treatment
effects (LATE). That is, the estimates measure the effects of grading on firms
whose graded absence share was manipulated by the grading practice of the
employees’ physicians. The effects on this group of compliers may be different
compared to effects of graded absence on “never-takers”, whether this is due to
the structure of production making it difficult to work when partially disabled
(e.g. physically and mentally demanding work, like firefighter or surgeon), or if
it is due to the underlying health condition causing the absences, i.e. diagnoses
that lead to complete loss of work function. Increasing the share of graded
absence in this group could conceivably have negative effects, as the disruption
to production would be unchanged while wage costs would increase.
For the instrumental variable strategy here to be valid, the following exclusion restriction should be satisfied: The physicians’ grading propensity should
influence firm profits only through the share of absence that is graded. Whether
or not this condition holds is in this case a matter of interpretation. For instance, Markussen et al, using a similar instrumental variable strategy, find that
graded absence certificates reduces the adjusted duration of the absence spell
and the likelihood of leaving the labor force, compared to full time absence.
This could also have consequences for the firm. In a way, this can be thought
of as one mechanism underlying the positive IV estimates of table 4: A higher
share of partial absence might improve profitability because partial absences
last a shorter time and fewer persons leave the firm for good, meaning there is
less hiring costs. The estimated effects should not be thought of as the effects
of simply changing the composition of absence, keeping all other things equal.
To supplement the IV estimates in table 4, table 5 reports estimates from
reduced form models that estimate the effects of physician practice style on
profits. As could be expected from the IV estimates, physician practice style
has no significant effects for the pooled sample of firms or for the low education
firms. For high education firms, a one percentage point increase in the average physician grading probability increases ROA by 0.000967. Relative to the
average annual ROA for high education firms in the sample (0.081), a one percentage point increase in the propensity of physicians to issue graded absence
certificates increases profitability by 1.2%.
A central assumption that should hold in order for the reported estimates to
have a causal interpretation is that any non-random sorting of firms to doctors
should be controlled for by the vector of observables xit . That is, the estimated
18
Table 5: Reduced Form, Firms
Grading propensity ×100
Observations
(1)
All
0.000224
(0.89)
152392
(2)
Low ed
-0.000133
(-0.34)
60595
(3)
High ed
0.000974∗∗
(2.02)
39427
t statistics in parentheses
∗ p < 0.1, ∗∗ p < 0.05, ∗∗∗ p < 0.01
Note: Models include controls for education background of the employees (share of employees
in each of 7 education levels), share of women, share of immigrants, average age, number of
employees, industry (68 categories), municipality and calendar year. Low education firms are
firms where 10% or less have a college degree; high education firms are firms where at least
33% of workers have a college degree. Standard error clustered at firm level.
effects of physician practice should reflect the effect of grading at the firm and
not uncontrolled for sorting. To check this, a robustness check is implemented
where the physician instrument at the firm level is calculated using only employees with no new sickness absence that started in the current year and the three
preceding years. For this group of workers, their absence patterns is unlikely
to have been influenced by their physician’s willingness to assign graded rather
than full time absence certificates. In the absence of sorting, there should then
be no effect of the instrument on firm profits.
Table 6: Placebo regression: never-absent workers
Grading propensity ×100
Observations
(1)
All
0.000142
(0.67)
140638
(2)
Low ed
-0.000133
(-0.43)
54612
(3)
High ed
0.000464
(1.05)
37208
t statistics in parentheses
∗ p < 0.1, ∗∗ p < 0.05, ∗∗∗ p < 0.01
Note: Placebo regression: Firm grading indicator constructed using only workers who did not
enter any sickness absence during the years t − 3 to t. Models include controls for education
background of the employees (share of employees in each of 7 education levels), share of women,
share of immigrants, average age, number of employees, industry (68 categories), municipality
and calendar year. Low education firms are firms where 10% or less have a college degree;
high education firms are firms where at least 33% of workers have a college degree. Standard
error clustered at firm level.
Results from this exercise are shown in table 6. The estimated effects of
grading propensity are not statistically significant for any firm category. While
the point estimates are have the same sign as the estimates in table 5, estimates
are closer to zero for all firms pooled together and for the high education firms
subsample. Overall, this exercise indicates that uncontrolled for sorting is not
19
driving the estimated effects of grading.
The estimates presented so far may be suffering from omitted variable bias if
physician grading practice is correlated with other aspects of physician practice
style. In particular, if doctors with high grading propensities are also ”stricter”
in the sense that they are less likely to issue absence certificates in the first place,
the estimated positive effects of grading propensity could be biased upward. Table 7 show estimates from extended models that include firm level indicators of
physician leniency. The leniency indicators included in these models reflect the
propensity of the employees’ physicians to issue absence certificates in general,
however, models where leniency indicators are based exclusively on issuance of
long term absence certificates yield similar results. As with the grading indicators, physician indicators are constructed in a way that controls for patient
characteristics including neighborhood of residence, demographics and work situation (see previous section for details).
Physician leniency is found to negatively impact firm profits (though the
estimates are not statistically significant). The estimated effects of grading
propensity are remarkably robust to the inclusion of these additional physician
characteristics.
Table 7: Extended models: physician leniency
Grading propensity ×100
Leniency indicator
Observations
(1)
All
0.000191
(0.75)
(2)
Low ed
-0.000171
(-0.44)
(3)
High ed
0.000953∗∗
(1.98)
-0.0825
(-1.54)
152391
-0.0755
(-0.93)
60594
-0.103
(-1.02)
39427
t statistics in parentheses
∗
p < 0.1,
∗∗
p < 0.05,
∗∗∗
p < 0.01
Note: Extended models. Models include controls for education background of the employees
(share of employees in each of 7 education levels), share of women, share of immigrants,
average age, number of employees, industry (68 categories), municipality and calendar year.
Low education firms are firms where 10% or less have a college degree; high education firms
are firms where at least 33% of workers have a college degree. Standard error clustered at
firm level.
So far, the estimated models have focused on impacts on the mean. The
impact of activation could be differ along the productivity distribution. If the
partial presence of a sick worker is most useful for firms that are close to failure, we would expect to see the largest impacts at the lower end of the firm
profitability distribution. However, it could also be that full time absence is the
most costly when times are good and there is a lot of work to do. In this case
we would expect the largest impacts on the top of the firm productivity distribution. To investigate this, I simultaneously estimate conditional quantiles at
20
the 25th, 50th and 75th percentile of the distribution of ROA.
Table 8: Quantile regressions
q25
Grading propensity ×100
q50
Grading propensity ×100
q75
Grading propensity ×100
Observations
(1)
All
(2)
Low ed
(3)
High ed
0.000209
(1.32)
-0.000257
(-1.03)
0.000674
(1.58)
0.000161
(0.71)
-0.000492∗
(-1.76)
0.000958∗∗∗
(2.84)
0.0000911
(0.32)
152393
-0.000398
(-0.89)
60595
0.00130∗∗∗
(3.11)
39427
t statistics in parentheses
∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01
Note: Models include controls for average age of workers, share of women and calendar year.
Low education firms are firms where 10% or less have a college degree; high education firms
are firms where at least 33% of workers have a college degree. Standard error clustered at
firm level.
Results are shown in table 8. For the full sample of firms, the estimated
effects of grading practice are positive across the distribution, however they are
(barely) significant only at the higher parts of the profit distribution. As before,
the impact of grading on low education firms is negative; for the median the
effect is significant at the ten percent level. For high education firms, the effects
of grading propensity on the 25th percentile are not statistically significant,
while effects are larger and significant for the 50th and 75th percentile. Overall,
the grading propensity of firms appears to be the most important for high
productivity, high education firms.
To summarize, the likelihood that a doctor issues a part time rather than a
full time absence certificate varies across individuals. Firms whose patients have
doctors with higher grading propensity have a higher share of partial absence.
Using this variation in graded absence rates, estimated models find that grading
appears to increase profits compared to full time absence. The effects are largest
for more knowledge intensive firms, defined as firms where a large share of
workers (> 33%) have a college education.
5
Conclusions
This paper asks the question of how firm profits are affected by activation requirements for workers on sickness absence. Descriptive statistics show a positive relationship between the share of long term absence that is graded and
21
the profits of the firm, as measured by the return on assets (ROA). This could
indicate that part time absence may be less costly for firms when the alternative is full time absence. Meanwhile, the use of graded rather than full time
absence certificates is likely to vary with observed and unobserved characteristics of firms. Data on use of graded sickness absence certificates indicate that
patients who are employed at high skill firms, defined as firms where a large
share of workers have a college education, are more likely to be issued a graded
absence certificate compared to people working at low skill firms.
An instrumental variable approach used to identify the causal effect of grading on firms finds that effects vary according to firm observable characteristics.
A strategy of issuing part time rather than full time absence certificates is found
to increase firm profits at high skill firms. The effects are largest and most significant at the top of the firm productivity distribution. For low skill firms, the
effects of grading are small and negative, though not statistically significant at
conventional significance levels.
Policymakers in comprehensive welfare states frequently face two potentially
contradictory policy goals: on the one hand, providing a generous welfare state
while on the other hand maintaining incentives for people to be economically
self-sufficient. The policy of encouraging physicians to issue graded rather than
full time absence certificates can be thought of as an attempt to reconcile these
goals (Markussen et al. 2012). Ideally, grading should reduce the costs to the
public of sickness absence, while still providing a high level of insurance to
workers with poor health.
This policy may be less desirable if grading simply shifts costs from taxpayers to firms. Grading could be costly to firms, through increased costs of
presenteeism (if workers are less productive than usual) or through direct costs
of accommodating partially disabled workers. Overall, the findings in this paper
indicate that such concerns are unwarranted. It would seem that the gains from
reduced interruptions of the production process generally outweigh the costs of
accommodating workers returning part time. This appears to be especially true
for high skill, highly productive firms, where grading has a net positive impact
on profitability.
References
Collins, J. J., Baase, C. M., Sharda, C. E., Ozminkowski, R. J., Nicholson, S.,
Billotti, G. M., Turpin, R. S., Olson, M. & Berger, M. L. (2005), ‘The assessment of chronic health conditions on work performance, absence, and total
economic impact for employers’, Journal of Occupational and Environmental
Medicine 47(6), 547–557.
Fevang, E., Markussen, S. & Røed, K. (2014), ‘The sick pay trap’, Journal of
Labor Economics 32(2), 305–336.
Høgelund, J., Holm, A. & McIntosh, J. (2010), ‘Does graded return-to-work
22
improve sick-listed workers’ chance of returning to regular working hours?’,
Journal of health economics 29(1), 158–169.
Markussen, S., Mykletun, A. & Røed, K. (2012), ‘The case for presenteeism—evidence from norway’s sickness insurance program’, Journal of Public Economics 96(11), 959–972.
Nicholson, S., Pauly, M. V., Polsky, D., Sharda, C., Szrek, H. & Berger, M. L.
(2006), ‘Measuring the effects of work loss on productivity with team production’, Health Economics 15(2), 111–123.
OECD (2010), Sickness, Disability and Work: Breaking the Barriers; A Synthesis of Findings Across OECD Countries, OECD.
Pauly, M. V., Nicholson, S., Xu, J., Polsky, D., Danzon, P. M., Murray, J. F.
& Berger, M. L. (2002), ‘A general model of the impact of absenteeism on
employers and employees’, Health Economics 11(3), 221–231.
Simon, G. E., Barber, C., Birnbaum, H. G., Frank, R. G., Greenberg, P. E.,
Rose, R. M., Wang, P. S. & Kessler, R. C. (2001), ‘Depression and work productivity: the comparative costs of treatment versus nontreatment’, Journal
of Occupational and Environmental Medicine 43(1), 2–9.
Stewart, W. F., Ricci, J. A., Chee, E., Hahn, S. R. & Morganstein, D. (2003),
‘Cost of lost productive work time among us workers with depression’, Jama
289(23), 3135–3144.
van Tulder, M. W., Koes, B. W. & Bouter, L. M. (1995), ‘A cost-of-illness study
of back pain in the netherlands’, Pain 62(2), 233–240.
Viikari-Juntura, E., Kausto, J., Shiri, R., Kaila-Kangas, L., Takala, E.-P., Karppinen, J., Miranda, H., Luukkonen, R. & Martimo, K.-P. (2012), ‘Return to
work after early part-time sick leave due to musculoskeletal disorders: a randomized controlled trial’, Scandinavian Journal of Work Environment and
Health 38(2), 134.
23