Manajemen | Fakultas Ekonomi Universitas Maritim Raja Ali Haji 156.full
Organizational Change, Absenteeism,
and Welfare Dependency
Knut Røed
Elisabeth Fevang
abstract
Based on Norwegian register data, we set up a multivariate mixed proportional hazard model (MMPH) to analyze nurses’ pattern of work, sickness
absence, nonemployment, and social insurance dependency from 1992 to
2000, and how that pattern was affected by workplace characteristics. The
model is estimated by means of the nonparametric maximum-likelihood estimator (NPMLE). We find that downsizing processes involve a significant
increase in the level of sickness absence among still-employed nurses. They
also cause a significant increase in the probability of entering into more
long-lasting health-related social insurance dependency.
I. Introduction
During the period from the early 1990s to 2003, Norway experienced
a remarkable increase in sickness absence and health-related withdrawals from the
labor market. The average number of sickness days paid for by the National Insurance Administration (NIA) increased by 75 percent, from a low of eight days per
worker in 1994 to 14 in 2003; see Figure 1, Panel A. And the rate at which new
Knut Røed is a senior research fellow and Elisabeth Fevang a research assistant at the Ragnar Frisch
Centre for Economic Research, Oslo, Norway. This paper is part of a Strategic Institute Program on Labor
Market and Pension Research, financed by the Norwegian Ministry of Labor and Social Inclusion and the
Norwegian Ministry of Finance. We have also received financial support from the Health Economics Research
Program at the University of Oslo (HERO). Thanks to Pål Jørgen Bakke, Morten Henningsen, Jonas
Lagerström, Karine Nyborg, Oddbjørn Raaum, and two anonymous referees, for valuable comments.
Also thanks to Morten Henningsen for assistance in the preparation of municipality/county data. The
administrative micro-data used in this paper are received from Statistics Norway by consent of the
Norwegian Data Inspectorate. The authors are not in a position to make the data available to other users,
but Statistics Norway can provide access subject to approval of an application. Interested Norwegian
scholars may consult www.ssb.no/mikrodata/. Access from outside Norway is restricted. Statistics Norway
may allow access, however, via statistical agencies in the country in question, provided they operate with
sufficient guarantees of confidentiality. The authors of this paper will help scholars seeking access to these
data as best they can. Correspondence to: Knut Røed, the Ragnar Frisch Centre for Economic Research,
Gaustadalléen 21, 0349 Oslo, Norway; www.frisch.uio.no. E-mail: [email protected].
[Submitted June 2005; accepted May 2006]
ISSN 022-166X E-ISSN 1548-8004 Ó 2007 by the Board of Regents of the University of Wisconsin System
T H E JO U R NAL O F H U M A N R ES O U R C ES
d
XLII
d
1
Sources: Absence data: the Norwegian National Insurance Administration (Rikstrygdeverket); unemployment data: OECD; turnover data: based on microdata
used in Dale-Olsen (2005) and kindly provided to us by the author; downsizing data: based on microdata collected by Statistics Norway and kindly provided to us
by Morten Henningsen (these data may be subject to future revision).
Røed and Fevang
Figure 1
Sickness absence, unemployment, turnover, and downsizing in Norway 1993–2003
157
158
The Journal of Human Resources
workers began absence spells that turned out to entail long-lasting social insurance
dependency (for more than six months) increased at a similar pace, from 2 percent
of the labor force in 1994 to just above 3 percent in 2000 (Fevang et al. 2004). Many
of these spells ended in permanent disability. Hence, the number of disability pensioners also rose sharply, from around 236,000 in the early 1990s to 301,000 by
the end of 2003.
Policymakers are concerned that this development is somehow related to the ‘‘brutalization’’ of work environments, with fiercer competition, faster structural change,
and an increased pressure to be adaptable. The pace of structural change has clearly
increased in sectors of the economy that were previously sheltered against competition, such as public utilities, telecommunications, postal services, healthcare, and
social services. There are also indications that the general levels of downsizing
and turnover have risen, see Figure 1, Panels C and D. In particular, firms’ downsizing propensity has increased sharply after 1995, despite general improvements in the
business cycle, see Panel B. The rise in turnover rates is a more predictable consequence of the cyclical developments. However, Dale-Olsen (2005) shows, on the basis of micro firm data, that even after controlling for business cycles (and changes in
industry composition) the turnover rate in Norwegian firms displayed a significant
positive trend during 1990–2002.1
A more turbulent labor market may have increased the risk of work-related health
impairments, and also made it more difficult for employees to be at work with given
health problems. As we show in the next section, some international evidence suggests that large organizational changes sometimes do have adverse health consequences. According to the influential demand-control theory (Karasek 1979; Karasek and
Theorell 1990), most work-related adverse reactions of psychological strain (fatigue,
anxiety, depression, and physical illness) occur when the psychological demands of
the job are high while the worker’s decision latitude in the task is low. It is conceivable that enforced organizational changes involving downsizing of the staff increase
the risk of such reactions. In Norway, as well as in other welfare state economies,
there has been a recent drive toward making the public sector more efficient, and
large organizational changes have been implemented to achieve this aim. Given
the strong employment protection in this sector, it is a worry that publicly financed
sickness and disability insurance (and ‘‘voluntary’’ withdrawals from the labor force)
sometimes serves as a substitute for layoffs.2
The present paper examines the relationship between organizational change, absenteeism, and welfare dependency within the public sector. It focuses on a group
of workers that faced substantial workplace turbulence during the 1990s—namely,
1. There is clearly also a business-cycle pattern in the sickness absence behaviour, but existing evidence
(Nordberg and Røed, 2003; Askildsen, Bratberg, and Nilsen 2005) indicates that this relationship is not very
strong. And business cycle effects can obviously not explain the rise in absenteeism that occurred after the
cyclical peak in 1998.
2. The sickness insurance system in Norway is universal and provides a 100 percent compensation rate for
one year and a substantially lower rehabilitation or disability benefit thereafter. Apart from the first 16 days
of absence (which are covered by the employer), the costs are paid for in full by the National Insurance
Administration (NIA). There is no experience rating.
Røed and Fevang
nurses working in hospitals, nursing homes, and community nursing.3 Our empirical
basis is the municipalities’ and counties’ own employment registers for 1992–2000.
These registers confirm that the pace of organizational change has increased somewhat in the healthcare and social services sector. Despite a rather strong overall employment growth (the total number of man-years in the sector rose by 23.8 percent
from 1992 to 2000), we find that the fraction of nurses who experienced an occurrence of large downsizing at their own workplace (encompassing at least 20 percent
of the man-years) increased from an average of around 2.1 percent per year in 1992–
95 to 2.9 percent in 1996–2000. An important distinguishing feature of downsizings
in this part of the economy is that they typically occur without layoffs. What usually
happens is that redundant (permanent contract) employees are offered alternative
employment opportunities within the municipality/county. Hence, the topic addressed in this paper is not the impact of individual displacement, but rather the consequences of experiencing a substantial organizational change, possibly involving
being transferred to a new workplace. Our key idea is to use observed organizational
changes, as they materialized in downsizing and/or large staff reshuffling, together
with longitudinal data on the employees’ individual work histories, to identify the
causal effects of organizational change on the workers’ absence behavior and subsequent employment and welfare dependency patterns.
Based on the population of employed nurses in October 1992, we set up a multivariate mixed proportional hazards (MMPH) model that facilitates an analysis of
how workplace events affect employee behavior in terms of transitions between presence and absence, between different workplaces, and between employment and nonemployment (with or without social insurance/assistance). To make this possible, we
have merged the employers’ employment registers with various public administrative
registers for the same period, covering all kinds of social security transfers to individuals (sickness benefits, unemployment benefits, social assistance, rehabilitation
benefits, and disability benefits). The statistical model we use imposes a minimum
of unjustified parametric restrictions. The various selection processes, including selection into the initial state, are modeled by means of the nonparametric maximum
likelihood estimator (NPMLE), which approximates the unknown distribution of unobserved heterogeneity in terms of a joint discrete distribution with an (a priori) unknown number of support points.
Our main finding is that major organizational changes do cause higher rates of absence and welfare dependency even when the changes do not involve layoffs. Nurses
employed at workplaces where large downsizing and/or substantial changes in the
composition of the staff occurred, experienced higher rates of long-term sickness
absence than nurses who worked in more stable economic environments, ceteris
3. In a recent sample survey among doctors and nurses in the Norwegian hospital sector (Sørensen and
Grimsmo 2004), as many as 59 percent of the respondents agreed to the statement that during the past
few years, considerations regarding the unit’s economic result had become a more important factor in their
daily work (compared with 4 percent who claimed it had become less important). A similar share of 54
percent felt that working speed had become a more important factor than before. In another study among
community nurses and home aids (Andersen and Eidset 2003), 57 percent of the respondents claimed that
strains resulting from time pressure had increased during the past two years (compared with 6 percent who
claimed it had been reduced). In comparison, only 24 percent felt that the physical strain had increased over
the same period.
159
160
The Journal of Human Resources
paribus. They also experienced higher rates of entry into other types of social security benefits as well as to self-supported nonemployment. In total, we find that a large
downsizing process (a reduction of at least 20 percent of the man-years during the
course of 12 months) causes between 1.5 and 3.0 percent of the nurses who worked
at the workplace in question initially to either become dependent of social insurance
benefits (primarily sickness, rehabilitation, or disability benefits) or to pull out of the
labor force without benefits. After the downsizing process is completed, the adverse
employment effect slowly peters out, but it takes more than six years before it is no
longer significantly different from zero.
The paper is structured as follows: The next section provides a brief review of the
existing literature. Section III describes the data. Section IV introduces the statistical
model. Section V presents the results, and Section VI concludes.
II. Existing Evidence
An extensive international economics literature covers individual
costs of displacement; for overviews, see, for example, Hamermesh (1987), Kletzer
(1998) and Kuhn (2002). The main conclusion arising from this literature is that the
costs associated with involuntary job loss are large and persistent in terms of both
subsequent income and employment propensity. A recent Norwegian study (Huttunen,
Møen, and Salvanes 2006) finds that displacement of male workers in the manufacturing industries substantially raises the probability of permanent exits from the labor
market. While 13 percent of the displaced workers in that study had left the labor
force seven years after the displacement occurred, this was the case for only 8 percent
of the control group containing similar nondisplaced workers. Another study (Rege,
Telle, and Votruba 2005) shows that workers employed in plants that closed between
1993 and 1998 were 28 percent more likely to be disability pensioners in 1999 than
comparable workers in nondownsizing plants. We have found no empirical evidence
in the economics literature regarding the impact of organizational changes on individuals who do not themselves become displaced.
However, a relatively large literature within epidemiological research suggests
that major organizational changes sometimes adversely affect employees’ health,
in terms of increased risks of cardiovascular diseases, musculoskeletal diseases, and
minor psychiatric morbidity; see for example Ferrie (2001) and Sverke, Hellgren, and
Näswall (2002) for recent surveys of the literature, and Burke and Greenglass (2000),
Burke (2003), Brown, Arentz, and Petterson (2003) and Sheward et al. (2005) for direct evidence regarding the healthcare and nursing sectors. These effects are not necessarily limited to increased job insecurity in the sense of layoffs, but may include
the loss of any valued condition of employment. Since data opportunities for identifying causal effects of this kind are rare due to the large selection problems involved,
most of the existing (reliable) research contributions have used a relatively small
number of ‘‘quasi-natural’’ experiments. The most famous of these is the Whitehall
II study (Ferrie et al. 1995; 1998a; 1998b; Ferrie et al. 2002). In 1988, a large reform
process was initiated in the British civil service, which for some departments involved a period of substantial uncertainty. One of the departments was later privatized. The evidence gathered from this ‘‘experiment’’ shows that the civil servants
Røed and Fevang
who faced job insecurity during this process experienced adverse health effects, both
in terms of self-rated health measures and clinical symptoms. Furthermore, the members of the department directly threatened by job loss due to privatization went from
a position of advantage or no difference in health status (prior to the ‘‘announcement’’ of the threat) to a significant disadvantage. Similar results are gathered from
a number of Finnish municipalities, which during the large recession in the 1990s
went through downsizing processes of varying degrees (Vahtera et al. 2000; Kivimäki
et al. 2000; Kivimäki et al. 2001; Vahtera et al. 2004; Vahtera et al. 2005). Downsizing turned out to be a risk factor for cardiovascular and musculoskeletal diseases
as well as for subsequent disability pension, even for individuals who remained in
employment.
Although there seems to be a negative causal relationship between job insecurity
and health, it is not necessarily the case that increased job insecurity also leads to
higher rates of sickness absence. The reason for this ambiguity is that job insecurity
also acts as a disciplinary device, such that more insecurity yields less absence for a
given health condition. According to the Finnish evidence, there is a significant positive association between the degree of downsizing and sickness absence among
permanent employees, but not among temporary employees (Vahtera et al. 2004).
Hence, if the threat of direct job loss is sufficiently large, the disciplinary effects
may neutralize or dominate the health effects. The Whitehall study also offers some
evidence in that direction. Individuals affected by general organizational changes had
a rise in sickness absence, whereas those facing imminent privatization of their department had a decline (Ferrie 2001).
There is some recent empirical evidence based on U.S. data that apparently does
not fit the picture of job insecurity adversely affecting health: Ruhm (2000; 2003)
shows that there is a significant countercyclical variation in physical health, and that
mortality rates correlate negatively with the rate of unemployment. However, aggregate economic fluctuations may affect health indicators differently than idiosyncratic
shocks, as stigma associated with being out of work diminishes during bad times, and
social networks to a larger extent move outside the workplaces; see, for example,
Kelvin and Jarret (1985). Moreover, economic downturns typically do not threaten
the job security of the majority of the population, hence adverse health effects for
those directly affected may be counteracted by lower workloads and healthier lifestyles for those who remain in secure employment (Ruhm 2005). There is also evidence that mental health as opposed to physical health varies procyclically (Ruhm
2003).
III. The Data: Employment, Welfare Dependency, and
Workplace Characteristics
The population under study in this paper comprises certified nurses
in Norway who were employed by a Norwegian municipality (home or community
nursing) or county (hospitals, psychiatric institutions), and who did not receive
any form of public income support (except parental leave or sickness benefits) by
the end of October 1992, and who were below 53 years of age at that time. The data
cover all municipalities and counties in Norway except the capital, Oslo, and include
161
162
The Journal of Human Resources
all permanent-contract employees.4 Starting in October 1992, these individuals are
monitored by the end of each month during the subsequent eight years. Descriptive
statistics are provided in Table 1. There are 43,167 individuals included in the analysis, of whom more than 95 percent are female. Half of the nurses are trained; that is,
they have at least three years of college education. The rest are enrolled nurses, with
vocational high-school education. Part I of Table 1 depicts the distribution of labor
market positions at the time of sampling. All individuals are employed at this point
(by construction of the data set), but the average number of contracted working hours
is only 74 percent of the full-time equivalent.5 Among young nurses there is also a
substantial absence due to parental leave. Absence due to long-term sickness (more
than 16 working days) is around 5 percent for the remaining trained nurses, and 7 percent for enrolled nurses. Part II of the table shows the distribution of labor market
states for the same group of individuals exactly eight years later. More than 80 percent of the nurses are still in self-supporting work and most of them remain employed
in healthcare or social services. There is, however, a sizable fraction of long-term
sickness absence in 2000 (9.2 percent), particularly among the nurses who stayed
on in these sectors. The absence level is much lower among nurses who switched
to another sector of the economy (5.3 percent). A point worth noting is that this apparently cannot be explained by nonrandom selection. On the contrary, nurses who
left had on average a higher level of sickness absence back in 1992 than those
who remained in the healthcare or social services sectors (not shown in the table).
As indicated in the last row of Table 1, 16.9 percent of the individuals in our sample received some kind of income support from the social security system by the end
of October 2000. Approximately half of them (8.8 percent) received a lasting benefit,
meaning a disability pension or a rehabilitation transfer. Layoffs seem to be rare
among nurses in the public sector, and only two per thousand nurses claimed unemployment benefits.
The overall rates of welfare dependency were much higher for enrolled nurses (21
percent) than for trained nurses (12 percent). Figure 2 provides a summary description of how these fractions evolved over time for our analysis population. As we have
conditioned on participation in the labor force to start with, and since the average age
of the analysis population increases with calendar time, it is of course not surprising
that these curves indicate rising social insurance dependency. The rate of increase
seems to be relatively stable over time, but with a strong seasonal pattern (the seasonality is primarily in sickness absence).
By merging the event histories of the individuals with other administrative registers, we have been able to extract information about the workplaces, the municipalities/counties responsible for running them, and the local labor market conditions;
see Table 2 for some selected descriptive statistics. In the remainder of this section,
we discuss some features of these data that are of interest with respect to identifying
the causal impact of organizational change on individuals’ absence and exit behavior.
4. Our data source distinguishes between short-term contracts (less than six months) and long-term contracts. We assume that the latter are permanent if they also contain a clearly defined regular working time
(to avoid ‘‘semi-permanent’’ substitute workers).
5. The number of actual working hours is higher because some nurses work more than contracted hours or
have more than one job.
Røed and Fevang
Table 1
Population under study. Descriptive statistics.
Trained Nurses
in 1992
All
Number of individuals
43,167
Mean age in 1992 (years)
38.0
Females (percent)
95.4
I. Status at the end of October 1992
Mean number of contracted working 74.2
hours relative to full time position
(percent)
Working in hospitals (percent)
42.0
Working in nursing homes (percent) 35.8
Working in psychiatric institutions
6.2
(percent)
Working in community nursing
16.0
(percent)
Absent due to parental leave (percent) 3.6
Fraction of remaining workers absent 6.1
due to long-term sickness (percent)a
II. Status at the end of October 2000
Still working without income support, 81.0
in the health-care and social service
sector (percent)
Out of which absent due to parental 0.8
leave (percent)
9.2
Fraction of remaining workers
absent due to long-term sickness
(percent)a
Still working without income support, 2.9
but in another sector of the
economy (percent)
Out of which absent due to parental 0.8
leave (percent)
5.3
Fraction of remaining workers
absent due to sickness (percent)a
Fully disabled (percent)
3.4
Partly disabled (percent)
2.5
Under rehabilitation (percent)
2.9
Enrolled Nurses
(Nurse’s aid)
in 1992
Age < 35 36–52 Age < 35 36–52
in 1992 in 1992 in 1992 in 1992
9,905
30.0
93.1
12,112
42.6
94.6
6,412
30.0
96.4
14,868
43.1
97.2
79.1
77.9
66.7
71.2
64.3
17.5
4.5
58.9
21.6
6.1
14.8
59.1
6.9
25.3
49.4
7.2
13.7
13.4
19.2
18.3
8.5
5.3
1.1
4.7
7.2
7.2
0.7
7.3
81.1
83.4
79.9
79.6
2.0
0.0
2.0
0.0
6.7
8.4
9.0
11.6
4.1
3.5
3.2
1.5
1.5
0.0
1.9
0.0
4.2
4.5
5.5
8.6
0.6
0.5
1.9
2.9
2.6
2.1
1.7
1.1
4.5
6.4
4.3
3.6
163
164
The Journal of Human Resources
Table 1 (continued )
Trained Nurses
in 1992
All
Unemployed (percent)
Living on means-tested social
assistance (percent)
Self-supported nonemployed
(percent)
Not registered (dead or migrated
to another country without
benefits) (percent)
Receiving some kind of social
security benefit (including
sickness benefits) (percent)
Enrolled Nurses
(Nurse’s aid)
in 1992
Age < 35 36–52 Age < 35 36–52
in 1992 in 1992 in 1992 in 1992
0.2
0.0
0.1
0.0
0.1
0.0
0.3
0.1
0.2
0.1
5.5
9.8
3.7
8.1
2.9
1.6
1.9
1.7
1.1
1.4
16.9
9.0
15.2
15.1
24.2
a. Long-term sickness is defined as having been absent from work for at least 16 working days. Shorter
absence spells are not recorded in administrative registers as they are not paid for by the National Insurance
Administration (but by the employer).
A full list of the explanatory variables that we actually use in the analysis is presented in the next section, after we have set up the statistical model.
Our key indicators for organizational change at workplaces are incidences of large
observed changes in the overall number of nursing man-years and (with some reservations explained below) large turnovers of the nursing staff. These variables are
both calculated from employment records in October each year. Let njs be the number of nurses at workplace j by the end of October in year s, and let njs be the corresponding total number of man-years (the full-time equivalent of njs ).6 Furthermore,
let injs be the number of new nurses in year s (those who were not employed at the
workplace in year s-1) and let outjs be the number of nurses who have left the workplace (those who were employed in year s-1, but not in s). We then define employnjs1 and turnover as minðinjs ; outjs Þ=njs1 .
ment growth in year s as ð
njs njs1 Þ=
(Turnover thus measures the fraction of employees that are ‘‘reshuffled,’’ that is,
the fraction of initial employees replaced by new ones.) Table 2 shows that the pace
of organizational change increased somewhat during the period considered in this paper, both in terms of downsizing and in terms of turnover. It also shows that this was
not primarily a result of general economic cutbacks. Both expenditures and incomes
increased more at the end of the data-period than in the beginning, and the unemployment rate (the definition of which is explained below) among all Norwegian
6. A ‘‘workplace’’ is typically a service institution, such as a hospital or a nursing home. Community
nursing is defined as one workplace for each of the 434 municipalities.
Røed and Fevang
Figure 2
Percent of nurses (employed in October 1992) receiving some kind of public
income support (including sickness benefits) October 1992–October 2000.
nurses fell from the already low level of 3.3 percent in 1992–95 to 1.6 percent in
1996–2000.
In the empirical analysis, we are particularly concerned with the impact of organizational changes that involve a substantive element of downsizing. Despite the relatively strong growth in the hospital and nursing sectors from 1992 to 2000 (2.7
percent growth in the number of man-years per year), Table 2 shows that a number
of workplaces reduced their staff. During the entire estimation period, 14 percent of
the nurses in our study experienced a downsizing of their workplace by more than 20
percent of the man-years, and 36 percent experienced a downsizing by more than 10
percent (these numbers are not shown in the table). Downsizings within nursing
homes and community nursing were typically linked to substantial reductions in
the municipality’s entire nursing staff, indicating that these organizational changes
were driven by the need to cut overall costs in the municipalities. Downsizings
within hospitals and psychiatric institutions on the other hand, were to a much larger
extent driven by reallocation of the jobs: in the hospital sector from small (and local)
to large hospitals, and in the psychiatry sector from institutions to outpatient clinics.
The combination of high employment growth and the strong employment protection
in this part of the economy implies that downsizings are typically implemented without layoffs. Instead, some employees are transferred to alternative workplaces within
the county/municipality. When this has to be done involuntarily, the seniority
165
166
The Journal of Human Resources
Table 2
The work environment: Selected characteristics of workplaces, employers and local
labor markets; by October Each Year
All
years 1992–95 1996–2000
Percent of nurses affected by the following
changes in the number of man-years
(at the workplace) from the previous year
>20 percent reduction
(10,20] percent reduction
(0,10] percent reduction
(0,10] percent increase
(10,20] percent increase
>20 percent increase
2.5
4.9
28.9
46.7
11.3
5.7
2.1
4.9
30.5
45.7
11.4
5.3
2.9
5.0
27.1
47.8
11.1
6.2
Percent of nurses affected by the following
turnover ratesa
15 percent
8.7
65.1
25.5
9.8
71.6
17.7
7.5
57.4
34.7
3.6
5.4
3.7
3.6
3.5
7.7
5.2
4.2
6.4
Mean debt-income ratio
5.3
Mean change in overall expenditures (from previous 6.2
year)
Mean change in overall income (from previous year) 4.6
Percent of employees in hospitals covered by the
25.6
pay-per-treatment regime
Mean local nurse unemployment rate
2.5
(mean taken over nurses)
4.8
5.3
5.7
7.3
3.6
0.0
5.8
64.9
3.3
1.6
County characteristics (mean taken over nurses)
Mean debt-income-ratio
Mean change in overall expenditures
(from previous year)
Mean change in overall income (from previous year)
Municipality characteristics (mean taken over nurses)
a. These numbers do not add up to 100 percent since turnover is not defined for employees in new establishments.
principle is normally applied, other things equal. The typical situation, however, is
that ‘‘other things’’ are not equal, as downsizings often involve transfers/closures
of whole units or the elimination of particular functions. Hence, it is difficult to identify the most affected workers, other than through actually observed transitions.
Some employees may also be pressured by the management to quit ‘‘voluntarily,’’
to claim sickness benefits, or to apply for disability pension. A possible side effect
Røed and Fevang
of the strong protection against layoffs is that health-related benefits to some extent
substitute for unemployment benefits.
While we consider changes in the overall number of nursing man-years to be exogenous with respect to the employment behavior of each nurse, this is obviously not the
case for the turnover variable (since each nurse’s quit behavior affects the turnover).
In the statistical model, we handle this problem by restricting our indicator of large
turnover to affect individual behavior with a time lag. Large turnover clearly does
not always reflect organizational changes. It may result from any source of employee
dissatisfaction and/or from particularly good alternative employment opportunities.
Organizational changes at a workplace do not necessarily lead to immediate and
visible changes in the staffing. In the hospital sector, for example, we suspect that
increased competitive pressures were generated by a reform in the hospital financing
system, from fully exogenous budgets to gradually increasing elements of the payper-treatment principle. The reform was introduced in July 1997 in 15 of Norway’s
19 counties (covering 75 percent of the hospital-based nurses in our data). The
remaining four counties followed suit, two of them as early as in January 1998 (raising the share of affected hospital nurses to 92 percent), one in January 1999 (raising
it to 97 percent), and the last one in January 2000. The purpose of the reform was
primarily to provide incentives for increased output (such as more surgical operations), not to cut public health spending. Hence, it typically did not result in downsizing. Because the reform was implemented in a staggered fashion, it is possible to
identify its causal effect on sickness absence and welfare dependency without relying
on calendar time or fixed county differences, provided that the staggered introduction
was exogenous. A possible reason to suspect endogenous introduction is that counties with highly efficient hospitals had stronger incentives to implement the reform
quickly than counties with less efficient hospitals. However, a recent study of technical efficiency in the Norwegian hospital sector found that there were no ex ante
differences in efficiency between the early and late implementers of the reform
(Biørn et al. 2003, p. 281). The main reason why some counties decided to wait
was that they wanted more discretionary control over medical priorities.
It is likely that many organizational changes that we are unable to observe have
been implemented throughout the municipalities and counties in Norway. We may
hypothesize, however, that the general pressure toward change has been stronger
where the local financial strain has been more severe. We have collected data from
the municipalities’/counties’ annual accounts, such as average income per inhabitant,
changes in gross yearly incomes, and the debt/income ratio. Unfortunately, these
indicators of economic vigor are not as good as we would have liked because we
do not observe the economic needs of each municipality/county, and, because it is
a stated aim of the income allocation system to provide equal standards across the
whole country, varying incomes to a large extent are likely to reflect varying needs.
We consider, however, a small yet interesting source of municipality income to be
exogenous, namely revenue generated from local natural resources (hydro power stations). Only a handful of municipalities have had the fortune of having these resources within their borders, but these municipalities are generally viewed as being
extremely rich. For historical reasons, they have succeeded in keeping these ‘‘windfall’’ gains out of the generally redistributive income system for Norwegian municipalities. The information that can be extracted from them in the present context is
167
168
The Journal of Human Resources
admittedly limited, however, as they are few and relatively small, and employ only
0.6 percent of the nurses in our study (around 240 persons).
Finally, we have collected (imperfect) information about the cost efficiency of the
municipalities’ nursing sector and their privatization strategies. The efficiency score
measures are collected from Erlandsen et al. (1997) and Edvardsen, Førsund, and
Aas (2000). They refer to two different points in time (1995 and 1997) and are based
on detailed accounts of costs and output in the nursing sector of each municipality.
The scores are computed by means of Data Envelopment Analysis (DEA), and, in
short, a municipality is considered to be fully efficient in the year under consideration if no convex combination of other municipalities exists that produces more care
services along all product dimension, without also using more of at least some inputs.
Our privatization indicator is collected from the so-called KOSTRA database (administered by Statistics Norway), and measures for each municipality in Norway
the fraction of total social service expenditures that were allocated through private
companies in 2003. Although both efficiency scores and privatization indicators
are measured with substantial errors (both because these concepts are difficult to
measure per se, and because they are measured for just a few years, one of them well
outside our observation window), we believe they may provide some information regarding the reform-strategies that were pursued during the 1990s, and hence the pace
of organizational change.
In addition to identifying the behavioral effects of organizational change, we also
wish to examine the possible impact of labor market tightness on the nurses’ employment and absence behavior. Because the nurses in our study typically enjoy very
strong employment protection, we expect discipline effects related to changes in
labor market tightness to be moderate, perhaps with an exception for nurses with
relatively small part time jobs (who might be underemployed). Unfortunately, the
standard local unemployment rates do not represent the employment opportunities
of nurses very well. We have therefore computed monthly county-specific unemployment rates for nurses in Norway particularly for the current project, based on the
whole population of certified nurses in Norway (not only the permanent contract
nurses covered in our analysis). These unemployment rates cover all kinds of unemployment registered at the employment offices.
IV. The Model: The Dynamics of Absence and Welfare
Dependency
In this section, we set up a competing risks hazard rate model aimed
at uncovering the causal effects of organizational change on the propensity to become and remain absent, and to become dependent on social insurance on a more
lasting basis. The model is set up in terms of monthly transition probabilities between
various labor market states, conditioned on vectors of observed and unobserved explanatory variables. Individuals can move among four mutually exclusive states:
(a) Present at work without reliance on social security benefits (this state also
includes absence spells shorter than 16 days, which are paid for by the
employer).
Røed and Fevang
(b) Absent from work with sickness benefits from the NIA.
(c) Nonemployed (or partly employed) with social security benefits (unemployment benefits, rehabilitation benefits, disability benefits, social assistance).
(d) Nonemployed without social security benefits.
In addition, individuals are allowed to move between different workplaces (given
that they are in State a or b). To start with (October 1992), all individuals belong to
one of the first two states; they are either present at work (State a), or absent due to
sickness (State b). Hence, the first part of our model is a binomial probability equation determining the initial state. In the subsequent months, transitions may occur.
Let t=1,2,.,97 be the 97 months for which we can (potentially) observe labor market states, where t=1 is the first month (October 1992). A transition is recorded whenever the state (or workplace) by the end of one month is different from the state (or
workplace) by the end of the previous month. In total, our state space allows for 13
different transition types (12 between the four states described above and one between different workplaces). However, in order to arrive at a tractable and economically interpretable model, we restrict some transitions with a common destination to
be governed by the same causal mechanisms regardless of the current state. In particular, we assume that both observed and unobserved covariates have the same effect on the hazard rates to the two nonemployment States (c and d) regardless of the
current state being presence (State a) or absence (State b). However, sickness absence is allowed to cause duration-dependent shifts in the levels of the hazard rates.
These effects are modeled by means of a timing-of-events approach (Abbring and
Van den Berg 2003a), implying that we do not need exclusion restrictions in order
to identify the causal effects of interest.7 In a similar fashion, we assume that observed and unobserved covariates have the same effect on the hazard rates from nonemployment and back to work presence, regardless of the current state being other
benefits (State c) or self-supported nonemployment (State d). However, the level
of the hazard rate may depend on whether the original exit involved a lasting benefit
or not. We do not model transitions between the two nonemployment States (c and
d); such transitions are only imperfectly observed and disregarded in the analysis.
Note also that we rule out transitions from the two nonemployment states directly
to sickness absence (from State c or d to State b). To the extent that such transitions
occur in the data, we interpret the sickness spell as a continuation of the nonemployment spell until work presence (State a) occurs. Based on these considerations, we
end up with the following distinct modeled transitions [the number of transitions
actually occurring in the data set in brackets]:
Transition 1: From work presence (State a) to long-term sickness absence (State b)
[73,493]
Transition 2: From long-term sickness absence (State b) to work presence (State a)
[61,450]
7. Note that according to Norwegian legislation, sickness absence cannot last longer than 12 months. At
this point some kind of transition must therefore take place.
169
170
The Journal of Human Resources
Figure 3
The modeled transitions (State a). The group ‘‘Not Employed’’ includes some
individuals who have become partly disabled, and have some (part-time)
employment (these are placed in the ‘‘other benefits’’ category).
Note: In addition to the six transitions illustrated in the figure, the initial condition (present or absent)
is also modeled as an endogenous event.
Transition 3: From a job at one workplace (State a or b) to another workplace
[20,683]
Transition 4: From a job (State a or b) to nonemployment benefits (State c) [7,486]
Transition 5: From a job (State a or b) to self-supported nonemployment (State d)
[6,617]
Transition 6: From nonemployment (State c or d) back to a job (State a) [7,576]
The transition pattern is illustrated in Figure 3. Individuals are included in the data
set as long as they do not make a transition to a job outside the workplaces covered
by our longitudinal data (in which case the spells are censored after the recording of
Transition 3). Spells are also exogenously censored for nurses who give birth to a
child, leave the country, or die. As and when these individuals return to the covered
workplaces, their spells are reactivated.
The hazard rates are assumed to depend on six groups of covariates; (i) calendar
time (t); (ii) individual and environmental characteristics (xit); (iii) municipality/
county characteristics (mit); (iv) workplace characteristics (zit); (v) duration of an ongoing sickness absence spell (dit); and (vi) unobserved individual (time-invariant)
characteristics (vi). The explanatory factors are allowed to vary between months
(except unobserved characteristics), but are assumed to be constant within months.
The hazard rates are specified within the framework of a multivariate mixed proportional hazard rate model (MMPH); see Abbring and Van den Berg (2003a; 2003b).
Røed and Fevang
Let uk ðwkit ; vi Þbe the hazard rate governing transitions to state k, given an index
value computed from observed covariates and model parameters wkit, and an unobserved (time-invariant) covariate vki. We specify the hazards in the following
way:
!
uk ðwkit ; vki Þ ¼ exp bk xit + rkt + lkd + ak mit + dk zit + vki ;
|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl}
wkit
ð1Þ
k ¼ 1; 2; :::; 6;
t ¼ 2; 3; :::; 97;
d ¼ 0; 1; :::; 12:
The exponential link-function in Equation 1 is used to ensure nonnegativity of the
hazards uk without constraining the permitted space for the parameters of interestðbk ; rkt ; lkd ; ak ; dk Þ. Note that since all explanatory variables and parameters
are assumed constant within each month, the hazards in Equation 1 can also be interpreted as integrated over these months. The parameters rkt and lkd are the calendar
time and duration effects, represented in the empirical model for transitions 1–6 by
separate dummy variables, one for each calendar month, and one for each possible
sickness absence duration. For transition 6 (from nonemployment and back to work),
more restrictive assumptions are required to estimate the model, given the much
smaller number of observations (the calendar time effects are assumed to be zero
and the duration effects are constrained to be piecewise constant in four-month periods). The initial condition is modeled in terms of a probability function l(.). Let y0i be
an indicator variable taking the value 1 if individual i was absent from work (due to
sickness) by the end of October 1992, and 0 otherwise. We then write the initial condition equation as
0
0
0
ð2Þ probðy0i ¼ 1Þ ¼ l xi1 b0 + m1 a0 + z1 d0 + v0i :
Let ykit 2 Y i be equal to 1 if month t involved a transition to state k for individual i,
and 0 otherwise, where Yi is the complete set of observed transition outcome indicators for individual i. Let K it be the set of feasible transitions for individual i in
month t; K it ¼ f2; 3; 4; 5g for individuals absent from work due to sickness,
K it ¼ f1; 3; 4; 5g for individuals present at work, and K it ¼ f6g for individuals without a job. We can then write the likelihood contribution formed by individual i, conditional on index functions for all relevant k and t ðwi Þand on the six-dimensional
vector of unobserved variables ðvi ¼ ðv0i ; v1i ; v2i ; v3i ; v4i ; v5i ; v6i ÞÞ as
ð3Þ
Li ðwi ; vi Þ ¼ P1 ðwi ; vi Þy0i ½1 P1 ðwi ; vi Þð1y0i Þ
Y Y
ykit 2Y i k2K it
½P2 ðwi ; vi ÞP3 ðwi ; vi Þykit ½1 P2 ðwi ; vi Þ
where
P1 ðwi ; vi Þ ¼ lðw0i ; v0i Þ
is the initial conditions probability Equation 2,
ð1 + ykit Þ
k2K it
;
171
172
The Journal of Human Resources
!
P2 ðwi ; vi Þ ¼ 1 exp + uk ðwkit ; vki Þ
k2K it
is the probability of making any of the available transitions in Kit, and
P3 ðwi ; vi Þ ¼
uk ðwkit ; vki Þ
+ uj ðwkit ; vki Þ
j2K it
is the conditional probability that the transition is of type k.8
Because the likelihood contribution in Equation 3 contains unobserved variables,
it cannot be used directly for estimation purposes. We approximate the unobserved
joint heterogeneity distribution in a nonparametric fashion with the aid of a discrete
distribution. Let Q be the (a priori unknown) number of support points in this distribution and let fvl ; pl g; l ¼ 1; 2; :::Q; be the associated location vectors and probabilities. In terms of observed variables, the likelihood function is then given as
ð4Þ
L¼
N
Y
i¼1
E½Li ðwi ; vi Þ ¼
vi
N
Y
Q
Q
+ pl Li ðwi ; vl Þ; + pl ¼ 1:
i¼1 l¼1
l¼1
To estimate the model, we also have to specify a functional form for the initial
condition probability function l(.) in Equation 2. For computational convenience,
we have chosen to specify it as a complementary log-log function, that is,
lð:Þ ¼ 1 expð expð:ÞÞ.9 Our estimation procedure is to maximize the likelihood
function in Equation 4 with respect to all the model and heterogeneity parameters
repeatedly for alternative values of Q. The nonparametric maximum likelihood estimators (NPMLE) are obtained by starting out with Q = 1, and then expanding
the model with new support points until the model is ‘‘saturated’’ in the sense that
it is no longer possible to increase the likelihood function by adding more points
(Lindsay 1983; Heckman and Singer 1984). At each stage of the estimation process,
we examine the appropriateness of an additional mass-point by means of simulated
annealing (Goffe, Ferrier, and Rogers 1994).10 Monte Carlo evidence presented in
Gaure, Røed, and Zhang (2005) indicates that there is little reason to believe that parameter estimates obtained this way are inconsistent or that they are nonnormally
8. The probability of making a transition to state k between (t-1) and t, given the feasible set of transitions
K (and given that hazard rates are assumed constant within each month), can then be derived as
Z t
Z u
Z t
uk ðwkit ; vki Þ exp +k2K
uk ðwkit ; vki Þds du ¼
uk ðwkit ; vki Þ exp +k2K uk ðwkit ; vki Þ
t1
t1
t1
u ðw ; v Þ
kit ki
k
:
3ðu ðt 1ÞÞ du ¼ 1 exp +k2K uk ðwkit ; vki Þ
+k uk2K ðwkit ; vki Þ
9. This simplifies the maximisation of the likelihood function, since it implies that the initial condition
obtains the same mathematical structure as the hazard-based elements of the likelihood. The complementary
log-log function has the unfortunate property that it is not symmetric. The results may therefore depend on
which outcome is defined as the base state. However, for the small event probabilities that are predominant
in our initial condition, the complementary log-log function is approximately equal to the logit-function.
10. Localisation of the nonparametric maximum-likelihood estimators is an enormous computational task,
and it could not have been done without extensive support from the High-Performance Computing Centre of
the University of Oslo. Our optimisation algorithm is described in detail in Gaure, Røed and Zhang (2005).
Røed and Fevang
distributed. They also indicate that the standard errors conditional on the optimal
number of support points are also valid for NPMLE.
A complete list of observed explanatory variables used in each of the seven equations (the initial condition and the six hazard rates) is provided in Table 3.11 Note that
changes in the level of workplace employment are allowed to have three different
types of effects on the various hazard rates in our model. First, current (ongoing)
changes in the employment level potentially affects all hazard rates; to sickness
absence (and back again), to other jobs, and to nonemployment (with or without
benefits).12 Second, some of these effects may persist after the changes are implemented. We allow these ‘‘lag-effects’’ to last for one year after implementation,
and also to differ between ‘‘survivors’’ (those who stay at the same workplace after
the changes) and ‘‘movers’’ (those who change to a new workplace, for example, as a
result of downsizing). Third, we may expect that the conditions (in terms of employment changes) under which some nurses leave employment (that is, make a transition
of Type 4 or 5) have a causal impact on their propensity to return to work. Note that
effects of downsizing on sickness absence behavior do not necessarily reflect effects
on individuals’ health; see Section II. They also may reflect effects on the threshold
at which a given health status is deemed to justify sickness absence and on the extent
to which sickness is untruthfully reported to conceal other reasons for absence.13 It is
not obvious whether s
and Welfare Dependency
Knut Røed
Elisabeth Fevang
abstract
Based on Norwegian register data, we set up a multivariate mixed proportional hazard model (MMPH) to analyze nurses’ pattern of work, sickness
absence, nonemployment, and social insurance dependency from 1992 to
2000, and how that pattern was affected by workplace characteristics. The
model is estimated by means of the nonparametric maximum-likelihood estimator (NPMLE). We find that downsizing processes involve a significant
increase in the level of sickness absence among still-employed nurses. They
also cause a significant increase in the probability of entering into more
long-lasting health-related social insurance dependency.
I. Introduction
During the period from the early 1990s to 2003, Norway experienced
a remarkable increase in sickness absence and health-related withdrawals from the
labor market. The average number of sickness days paid for by the National Insurance Administration (NIA) increased by 75 percent, from a low of eight days per
worker in 1994 to 14 in 2003; see Figure 1, Panel A. And the rate at which new
Knut Røed is a senior research fellow and Elisabeth Fevang a research assistant at the Ragnar Frisch
Centre for Economic Research, Oslo, Norway. This paper is part of a Strategic Institute Program on Labor
Market and Pension Research, financed by the Norwegian Ministry of Labor and Social Inclusion and the
Norwegian Ministry of Finance. We have also received financial support from the Health Economics Research
Program at the University of Oslo (HERO). Thanks to Pål Jørgen Bakke, Morten Henningsen, Jonas
Lagerström, Karine Nyborg, Oddbjørn Raaum, and two anonymous referees, for valuable comments.
Also thanks to Morten Henningsen for assistance in the preparation of municipality/county data. The
administrative micro-data used in this paper are received from Statistics Norway by consent of the
Norwegian Data Inspectorate. The authors are not in a position to make the data available to other users,
but Statistics Norway can provide access subject to approval of an application. Interested Norwegian
scholars may consult www.ssb.no/mikrodata/. Access from outside Norway is restricted. Statistics Norway
may allow access, however, via statistical agencies in the country in question, provided they operate with
sufficient guarantees of confidentiality. The authors of this paper will help scholars seeking access to these
data as best they can. Correspondence to: Knut Røed, the Ragnar Frisch Centre for Economic Research,
Gaustadalléen 21, 0349 Oslo, Norway; www.frisch.uio.no. E-mail: [email protected].
[Submitted June 2005; accepted May 2006]
ISSN 022-166X E-ISSN 1548-8004 Ó 2007 by the Board of Regents of the University of Wisconsin System
T H E JO U R NAL O F H U M A N R ES O U R C ES
d
XLII
d
1
Sources: Absence data: the Norwegian National Insurance Administration (Rikstrygdeverket); unemployment data: OECD; turnover data: based on microdata
used in Dale-Olsen (2005) and kindly provided to us by the author; downsizing data: based on microdata collected by Statistics Norway and kindly provided to us
by Morten Henningsen (these data may be subject to future revision).
Røed and Fevang
Figure 1
Sickness absence, unemployment, turnover, and downsizing in Norway 1993–2003
157
158
The Journal of Human Resources
workers began absence spells that turned out to entail long-lasting social insurance
dependency (for more than six months) increased at a similar pace, from 2 percent
of the labor force in 1994 to just above 3 percent in 2000 (Fevang et al. 2004). Many
of these spells ended in permanent disability. Hence, the number of disability pensioners also rose sharply, from around 236,000 in the early 1990s to 301,000 by
the end of 2003.
Policymakers are concerned that this development is somehow related to the ‘‘brutalization’’ of work environments, with fiercer competition, faster structural change,
and an increased pressure to be adaptable. The pace of structural change has clearly
increased in sectors of the economy that were previously sheltered against competition, such as public utilities, telecommunications, postal services, healthcare, and
social services. There are also indications that the general levels of downsizing
and turnover have risen, see Figure 1, Panels C and D. In particular, firms’ downsizing propensity has increased sharply after 1995, despite general improvements in the
business cycle, see Panel B. The rise in turnover rates is a more predictable consequence of the cyclical developments. However, Dale-Olsen (2005) shows, on the basis of micro firm data, that even after controlling for business cycles (and changes in
industry composition) the turnover rate in Norwegian firms displayed a significant
positive trend during 1990–2002.1
A more turbulent labor market may have increased the risk of work-related health
impairments, and also made it more difficult for employees to be at work with given
health problems. As we show in the next section, some international evidence suggests that large organizational changes sometimes do have adverse health consequences. According to the influential demand-control theory (Karasek 1979; Karasek and
Theorell 1990), most work-related adverse reactions of psychological strain (fatigue,
anxiety, depression, and physical illness) occur when the psychological demands of
the job are high while the worker’s decision latitude in the task is low. It is conceivable that enforced organizational changes involving downsizing of the staff increase
the risk of such reactions. In Norway, as well as in other welfare state economies,
there has been a recent drive toward making the public sector more efficient, and
large organizational changes have been implemented to achieve this aim. Given
the strong employment protection in this sector, it is a worry that publicly financed
sickness and disability insurance (and ‘‘voluntary’’ withdrawals from the labor force)
sometimes serves as a substitute for layoffs.2
The present paper examines the relationship between organizational change, absenteeism, and welfare dependency within the public sector. It focuses on a group
of workers that faced substantial workplace turbulence during the 1990s—namely,
1. There is clearly also a business-cycle pattern in the sickness absence behaviour, but existing evidence
(Nordberg and Røed, 2003; Askildsen, Bratberg, and Nilsen 2005) indicates that this relationship is not very
strong. And business cycle effects can obviously not explain the rise in absenteeism that occurred after the
cyclical peak in 1998.
2. The sickness insurance system in Norway is universal and provides a 100 percent compensation rate for
one year and a substantially lower rehabilitation or disability benefit thereafter. Apart from the first 16 days
of absence (which are covered by the employer), the costs are paid for in full by the National Insurance
Administration (NIA). There is no experience rating.
Røed and Fevang
nurses working in hospitals, nursing homes, and community nursing.3 Our empirical
basis is the municipalities’ and counties’ own employment registers for 1992–2000.
These registers confirm that the pace of organizational change has increased somewhat in the healthcare and social services sector. Despite a rather strong overall employment growth (the total number of man-years in the sector rose by 23.8 percent
from 1992 to 2000), we find that the fraction of nurses who experienced an occurrence of large downsizing at their own workplace (encompassing at least 20 percent
of the man-years) increased from an average of around 2.1 percent per year in 1992–
95 to 2.9 percent in 1996–2000. An important distinguishing feature of downsizings
in this part of the economy is that they typically occur without layoffs. What usually
happens is that redundant (permanent contract) employees are offered alternative
employment opportunities within the municipality/county. Hence, the topic addressed in this paper is not the impact of individual displacement, but rather the consequences of experiencing a substantial organizational change, possibly involving
being transferred to a new workplace. Our key idea is to use observed organizational
changes, as they materialized in downsizing and/or large staff reshuffling, together
with longitudinal data on the employees’ individual work histories, to identify the
causal effects of organizational change on the workers’ absence behavior and subsequent employment and welfare dependency patterns.
Based on the population of employed nurses in October 1992, we set up a multivariate mixed proportional hazards (MMPH) model that facilitates an analysis of
how workplace events affect employee behavior in terms of transitions between presence and absence, between different workplaces, and between employment and nonemployment (with or without social insurance/assistance). To make this possible, we
have merged the employers’ employment registers with various public administrative
registers for the same period, covering all kinds of social security transfers to individuals (sickness benefits, unemployment benefits, social assistance, rehabilitation
benefits, and disability benefits). The statistical model we use imposes a minimum
of unjustified parametric restrictions. The various selection processes, including selection into the initial state, are modeled by means of the nonparametric maximum
likelihood estimator (NPMLE), which approximates the unknown distribution of unobserved heterogeneity in terms of a joint discrete distribution with an (a priori) unknown number of support points.
Our main finding is that major organizational changes do cause higher rates of absence and welfare dependency even when the changes do not involve layoffs. Nurses
employed at workplaces where large downsizing and/or substantial changes in the
composition of the staff occurred, experienced higher rates of long-term sickness
absence than nurses who worked in more stable economic environments, ceteris
3. In a recent sample survey among doctors and nurses in the Norwegian hospital sector (Sørensen and
Grimsmo 2004), as many as 59 percent of the respondents agreed to the statement that during the past
few years, considerations regarding the unit’s economic result had become a more important factor in their
daily work (compared with 4 percent who claimed it had become less important). A similar share of 54
percent felt that working speed had become a more important factor than before. In another study among
community nurses and home aids (Andersen and Eidset 2003), 57 percent of the respondents claimed that
strains resulting from time pressure had increased during the past two years (compared with 6 percent who
claimed it had been reduced). In comparison, only 24 percent felt that the physical strain had increased over
the same period.
159
160
The Journal of Human Resources
paribus. They also experienced higher rates of entry into other types of social security benefits as well as to self-supported nonemployment. In total, we find that a large
downsizing process (a reduction of at least 20 percent of the man-years during the
course of 12 months) causes between 1.5 and 3.0 percent of the nurses who worked
at the workplace in question initially to either become dependent of social insurance
benefits (primarily sickness, rehabilitation, or disability benefits) or to pull out of the
labor force without benefits. After the downsizing process is completed, the adverse
employment effect slowly peters out, but it takes more than six years before it is no
longer significantly different from zero.
The paper is structured as follows: The next section provides a brief review of the
existing literature. Section III describes the data. Section IV introduces the statistical
model. Section V presents the results, and Section VI concludes.
II. Existing Evidence
An extensive international economics literature covers individual
costs of displacement; for overviews, see, for example, Hamermesh (1987), Kletzer
(1998) and Kuhn (2002). The main conclusion arising from this literature is that the
costs associated with involuntary job loss are large and persistent in terms of both
subsequent income and employment propensity. A recent Norwegian study (Huttunen,
Møen, and Salvanes 2006) finds that displacement of male workers in the manufacturing industries substantially raises the probability of permanent exits from the labor
market. While 13 percent of the displaced workers in that study had left the labor
force seven years after the displacement occurred, this was the case for only 8 percent
of the control group containing similar nondisplaced workers. Another study (Rege,
Telle, and Votruba 2005) shows that workers employed in plants that closed between
1993 and 1998 were 28 percent more likely to be disability pensioners in 1999 than
comparable workers in nondownsizing plants. We have found no empirical evidence
in the economics literature regarding the impact of organizational changes on individuals who do not themselves become displaced.
However, a relatively large literature within epidemiological research suggests
that major organizational changes sometimes adversely affect employees’ health,
in terms of increased risks of cardiovascular diseases, musculoskeletal diseases, and
minor psychiatric morbidity; see for example Ferrie (2001) and Sverke, Hellgren, and
Näswall (2002) for recent surveys of the literature, and Burke and Greenglass (2000),
Burke (2003), Brown, Arentz, and Petterson (2003) and Sheward et al. (2005) for direct evidence regarding the healthcare and nursing sectors. These effects are not necessarily limited to increased job insecurity in the sense of layoffs, but may include
the loss of any valued condition of employment. Since data opportunities for identifying causal effects of this kind are rare due to the large selection problems involved,
most of the existing (reliable) research contributions have used a relatively small
number of ‘‘quasi-natural’’ experiments. The most famous of these is the Whitehall
II study (Ferrie et al. 1995; 1998a; 1998b; Ferrie et al. 2002). In 1988, a large reform
process was initiated in the British civil service, which for some departments involved a period of substantial uncertainty. One of the departments was later privatized. The evidence gathered from this ‘‘experiment’’ shows that the civil servants
Røed and Fevang
who faced job insecurity during this process experienced adverse health effects, both
in terms of self-rated health measures and clinical symptoms. Furthermore, the members of the department directly threatened by job loss due to privatization went from
a position of advantage or no difference in health status (prior to the ‘‘announcement’’ of the threat) to a significant disadvantage. Similar results are gathered from
a number of Finnish municipalities, which during the large recession in the 1990s
went through downsizing processes of varying degrees (Vahtera et al. 2000; Kivimäki
et al. 2000; Kivimäki et al. 2001; Vahtera et al. 2004; Vahtera et al. 2005). Downsizing turned out to be a risk factor for cardiovascular and musculoskeletal diseases
as well as for subsequent disability pension, even for individuals who remained in
employment.
Although there seems to be a negative causal relationship between job insecurity
and health, it is not necessarily the case that increased job insecurity also leads to
higher rates of sickness absence. The reason for this ambiguity is that job insecurity
also acts as a disciplinary device, such that more insecurity yields less absence for a
given health condition. According to the Finnish evidence, there is a significant positive association between the degree of downsizing and sickness absence among
permanent employees, but not among temporary employees (Vahtera et al. 2004).
Hence, if the threat of direct job loss is sufficiently large, the disciplinary effects
may neutralize or dominate the health effects. The Whitehall study also offers some
evidence in that direction. Individuals affected by general organizational changes had
a rise in sickness absence, whereas those facing imminent privatization of their department had a decline (Ferrie 2001).
There is some recent empirical evidence based on U.S. data that apparently does
not fit the picture of job insecurity adversely affecting health: Ruhm (2000; 2003)
shows that there is a significant countercyclical variation in physical health, and that
mortality rates correlate negatively with the rate of unemployment. However, aggregate economic fluctuations may affect health indicators differently than idiosyncratic
shocks, as stigma associated with being out of work diminishes during bad times, and
social networks to a larger extent move outside the workplaces; see, for example,
Kelvin and Jarret (1985). Moreover, economic downturns typically do not threaten
the job security of the majority of the population, hence adverse health effects for
those directly affected may be counteracted by lower workloads and healthier lifestyles for those who remain in secure employment (Ruhm 2005). There is also evidence that mental health as opposed to physical health varies procyclically (Ruhm
2003).
III. The Data: Employment, Welfare Dependency, and
Workplace Characteristics
The population under study in this paper comprises certified nurses
in Norway who were employed by a Norwegian municipality (home or community
nursing) or county (hospitals, psychiatric institutions), and who did not receive
any form of public income support (except parental leave or sickness benefits) by
the end of October 1992, and who were below 53 years of age at that time. The data
cover all municipalities and counties in Norway except the capital, Oslo, and include
161
162
The Journal of Human Resources
all permanent-contract employees.4 Starting in October 1992, these individuals are
monitored by the end of each month during the subsequent eight years. Descriptive
statistics are provided in Table 1. There are 43,167 individuals included in the analysis, of whom more than 95 percent are female. Half of the nurses are trained; that is,
they have at least three years of college education. The rest are enrolled nurses, with
vocational high-school education. Part I of Table 1 depicts the distribution of labor
market positions at the time of sampling. All individuals are employed at this point
(by construction of the data set), but the average number of contracted working hours
is only 74 percent of the full-time equivalent.5 Among young nurses there is also a
substantial absence due to parental leave. Absence due to long-term sickness (more
than 16 working days) is around 5 percent for the remaining trained nurses, and 7 percent for enrolled nurses. Part II of the table shows the distribution of labor market
states for the same group of individuals exactly eight years later. More than 80 percent of the nurses are still in self-supporting work and most of them remain employed
in healthcare or social services. There is, however, a sizable fraction of long-term
sickness absence in 2000 (9.2 percent), particularly among the nurses who stayed
on in these sectors. The absence level is much lower among nurses who switched
to another sector of the economy (5.3 percent). A point worth noting is that this apparently cannot be explained by nonrandom selection. On the contrary, nurses who
left had on average a higher level of sickness absence back in 1992 than those
who remained in the healthcare or social services sectors (not shown in the table).
As indicated in the last row of Table 1, 16.9 percent of the individuals in our sample received some kind of income support from the social security system by the end
of October 2000. Approximately half of them (8.8 percent) received a lasting benefit,
meaning a disability pension or a rehabilitation transfer. Layoffs seem to be rare
among nurses in the public sector, and only two per thousand nurses claimed unemployment benefits.
The overall rates of welfare dependency were much higher for enrolled nurses (21
percent) than for trained nurses (12 percent). Figure 2 provides a summary description of how these fractions evolved over time for our analysis population. As we have
conditioned on participation in the labor force to start with, and since the average age
of the analysis population increases with calendar time, it is of course not surprising
that these curves indicate rising social insurance dependency. The rate of increase
seems to be relatively stable over time, but with a strong seasonal pattern (the seasonality is primarily in sickness absence).
By merging the event histories of the individuals with other administrative registers, we have been able to extract information about the workplaces, the municipalities/counties responsible for running them, and the local labor market conditions;
see Table 2 for some selected descriptive statistics. In the remainder of this section,
we discuss some features of these data that are of interest with respect to identifying
the causal impact of organizational change on individuals’ absence and exit behavior.
4. Our data source distinguishes between short-term contracts (less than six months) and long-term contracts. We assume that the latter are permanent if they also contain a clearly defined regular working time
(to avoid ‘‘semi-permanent’’ substitute workers).
5. The number of actual working hours is higher because some nurses work more than contracted hours or
have more than one job.
Røed and Fevang
Table 1
Population under study. Descriptive statistics.
Trained Nurses
in 1992
All
Number of individuals
43,167
Mean age in 1992 (years)
38.0
Females (percent)
95.4
I. Status at the end of October 1992
Mean number of contracted working 74.2
hours relative to full time position
(percent)
Working in hospitals (percent)
42.0
Working in nursing homes (percent) 35.8
Working in psychiatric institutions
6.2
(percent)
Working in community nursing
16.0
(percent)
Absent due to parental leave (percent) 3.6
Fraction of remaining workers absent 6.1
due to long-term sickness (percent)a
II. Status at the end of October 2000
Still working without income support, 81.0
in the health-care and social service
sector (percent)
Out of which absent due to parental 0.8
leave (percent)
9.2
Fraction of remaining workers
absent due to long-term sickness
(percent)a
Still working without income support, 2.9
but in another sector of the
economy (percent)
Out of which absent due to parental 0.8
leave (percent)
5.3
Fraction of remaining workers
absent due to sickness (percent)a
Fully disabled (percent)
3.4
Partly disabled (percent)
2.5
Under rehabilitation (percent)
2.9
Enrolled Nurses
(Nurse’s aid)
in 1992
Age < 35 36–52 Age < 35 36–52
in 1992 in 1992 in 1992 in 1992
9,905
30.0
93.1
12,112
42.6
94.6
6,412
30.0
96.4
14,868
43.1
97.2
79.1
77.9
66.7
71.2
64.3
17.5
4.5
58.9
21.6
6.1
14.8
59.1
6.9
25.3
49.4
7.2
13.7
13.4
19.2
18.3
8.5
5.3
1.1
4.7
7.2
7.2
0.7
7.3
81.1
83.4
79.9
79.6
2.0
0.0
2.0
0.0
6.7
8.4
9.0
11.6
4.1
3.5
3.2
1.5
1.5
0.0
1.9
0.0
4.2
4.5
5.5
8.6
0.6
0.5
1.9
2.9
2.6
2.1
1.7
1.1
4.5
6.4
4.3
3.6
163
164
The Journal of Human Resources
Table 1 (continued )
Trained Nurses
in 1992
All
Unemployed (percent)
Living on means-tested social
assistance (percent)
Self-supported nonemployed
(percent)
Not registered (dead or migrated
to another country without
benefits) (percent)
Receiving some kind of social
security benefit (including
sickness benefits) (percent)
Enrolled Nurses
(Nurse’s aid)
in 1992
Age < 35 36–52 Age < 35 36–52
in 1992 in 1992 in 1992 in 1992
0.2
0.0
0.1
0.0
0.1
0.0
0.3
0.1
0.2
0.1
5.5
9.8
3.7
8.1
2.9
1.6
1.9
1.7
1.1
1.4
16.9
9.0
15.2
15.1
24.2
a. Long-term sickness is defined as having been absent from work for at least 16 working days. Shorter
absence spells are not recorded in administrative registers as they are not paid for by the National Insurance
Administration (but by the employer).
A full list of the explanatory variables that we actually use in the analysis is presented in the next section, after we have set up the statistical model.
Our key indicators for organizational change at workplaces are incidences of large
observed changes in the overall number of nursing man-years and (with some reservations explained below) large turnovers of the nursing staff. These variables are
both calculated from employment records in October each year. Let njs be the number of nurses at workplace j by the end of October in year s, and let njs be the corresponding total number of man-years (the full-time equivalent of njs ).6 Furthermore,
let injs be the number of new nurses in year s (those who were not employed at the
workplace in year s-1) and let outjs be the number of nurses who have left the workplace (those who were employed in year s-1, but not in s). We then define employnjs1 and turnover as minðinjs ; outjs Þ=njs1 .
ment growth in year s as ð
njs njs1 Þ=
(Turnover thus measures the fraction of employees that are ‘‘reshuffled,’’ that is,
the fraction of initial employees replaced by new ones.) Table 2 shows that the pace
of organizational change increased somewhat during the period considered in this paper, both in terms of downsizing and in terms of turnover. It also shows that this was
not primarily a result of general economic cutbacks. Both expenditures and incomes
increased more at the end of the data-period than in the beginning, and the unemployment rate (the definition of which is explained below) among all Norwegian
6. A ‘‘workplace’’ is typically a service institution, such as a hospital or a nursing home. Community
nursing is defined as one workplace for each of the 434 municipalities.
Røed and Fevang
Figure 2
Percent of nurses (employed in October 1992) receiving some kind of public
income support (including sickness benefits) October 1992–October 2000.
nurses fell from the already low level of 3.3 percent in 1992–95 to 1.6 percent in
1996–2000.
In the empirical analysis, we are particularly concerned with the impact of organizational changes that involve a substantive element of downsizing. Despite the relatively strong growth in the hospital and nursing sectors from 1992 to 2000 (2.7
percent growth in the number of man-years per year), Table 2 shows that a number
of workplaces reduced their staff. During the entire estimation period, 14 percent of
the nurses in our study experienced a downsizing of their workplace by more than 20
percent of the man-years, and 36 percent experienced a downsizing by more than 10
percent (these numbers are not shown in the table). Downsizings within nursing
homes and community nursing were typically linked to substantial reductions in
the municipality’s entire nursing staff, indicating that these organizational changes
were driven by the need to cut overall costs in the municipalities. Downsizings
within hospitals and psychiatric institutions on the other hand, were to a much larger
extent driven by reallocation of the jobs: in the hospital sector from small (and local)
to large hospitals, and in the psychiatry sector from institutions to outpatient clinics.
The combination of high employment growth and the strong employment protection
in this part of the economy implies that downsizings are typically implemented without layoffs. Instead, some employees are transferred to alternative workplaces within
the county/municipality. When this has to be done involuntarily, the seniority
165
166
The Journal of Human Resources
Table 2
The work environment: Selected characteristics of workplaces, employers and local
labor markets; by October Each Year
All
years 1992–95 1996–2000
Percent of nurses affected by the following
changes in the number of man-years
(at the workplace) from the previous year
>20 percent reduction
(10,20] percent reduction
(0,10] percent reduction
(0,10] percent increase
(10,20] percent increase
>20 percent increase
2.5
4.9
28.9
46.7
11.3
5.7
2.1
4.9
30.5
45.7
11.4
5.3
2.9
5.0
27.1
47.8
11.1
6.2
Percent of nurses affected by the following
turnover ratesa
15 percent
8.7
65.1
25.5
9.8
71.6
17.7
7.5
57.4
34.7
3.6
5.4
3.7
3.6
3.5
7.7
5.2
4.2
6.4
Mean debt-income ratio
5.3
Mean change in overall expenditures (from previous 6.2
year)
Mean change in overall income (from previous year) 4.6
Percent of employees in hospitals covered by the
25.6
pay-per-treatment regime
Mean local nurse unemployment rate
2.5
(mean taken over nurses)
4.8
5.3
5.7
7.3
3.6
0.0
5.8
64.9
3.3
1.6
County characteristics (mean taken over nurses)
Mean debt-income-ratio
Mean change in overall expenditures
(from previous year)
Mean change in overall income (from previous year)
Municipality characteristics (mean taken over nurses)
a. These numbers do not add up to 100 percent since turnover is not defined for employees in new establishments.
principle is normally applied, other things equal. The typical situation, however, is
that ‘‘other things’’ are not equal, as downsizings often involve transfers/closures
of whole units or the elimination of particular functions. Hence, it is difficult to identify the most affected workers, other than through actually observed transitions.
Some employees may also be pressured by the management to quit ‘‘voluntarily,’’
to claim sickness benefits, or to apply for disability pension. A possible side effect
Røed and Fevang
of the strong protection against layoffs is that health-related benefits to some extent
substitute for unemployment benefits.
While we consider changes in the overall number of nursing man-years to be exogenous with respect to the employment behavior of each nurse, this is obviously not the
case for the turnover variable (since each nurse’s quit behavior affects the turnover).
In the statistical model, we handle this problem by restricting our indicator of large
turnover to affect individual behavior with a time lag. Large turnover clearly does
not always reflect organizational changes. It may result from any source of employee
dissatisfaction and/or from particularly good alternative employment opportunities.
Organizational changes at a workplace do not necessarily lead to immediate and
visible changes in the staffing. In the hospital sector, for example, we suspect that
increased competitive pressures were generated by a reform in the hospital financing
system, from fully exogenous budgets to gradually increasing elements of the payper-treatment principle. The reform was introduced in July 1997 in 15 of Norway’s
19 counties (covering 75 percent of the hospital-based nurses in our data). The
remaining four counties followed suit, two of them as early as in January 1998 (raising the share of affected hospital nurses to 92 percent), one in January 1999 (raising
it to 97 percent), and the last one in January 2000. The purpose of the reform was
primarily to provide incentives for increased output (such as more surgical operations), not to cut public health spending. Hence, it typically did not result in downsizing. Because the reform was implemented in a staggered fashion, it is possible to
identify its causal effect on sickness absence and welfare dependency without relying
on calendar time or fixed county differences, provided that the staggered introduction
was exogenous. A possible reason to suspect endogenous introduction is that counties with highly efficient hospitals had stronger incentives to implement the reform
quickly than counties with less efficient hospitals. However, a recent study of technical efficiency in the Norwegian hospital sector found that there were no ex ante
differences in efficiency between the early and late implementers of the reform
(Biørn et al. 2003, p. 281). The main reason why some counties decided to wait
was that they wanted more discretionary control over medical priorities.
It is likely that many organizational changes that we are unable to observe have
been implemented throughout the municipalities and counties in Norway. We may
hypothesize, however, that the general pressure toward change has been stronger
where the local financial strain has been more severe. We have collected data from
the municipalities’/counties’ annual accounts, such as average income per inhabitant,
changes in gross yearly incomes, and the debt/income ratio. Unfortunately, these
indicators of economic vigor are not as good as we would have liked because we
do not observe the economic needs of each municipality/county, and, because it is
a stated aim of the income allocation system to provide equal standards across the
whole country, varying incomes to a large extent are likely to reflect varying needs.
We consider, however, a small yet interesting source of municipality income to be
exogenous, namely revenue generated from local natural resources (hydro power stations). Only a handful of municipalities have had the fortune of having these resources within their borders, but these municipalities are generally viewed as being
extremely rich. For historical reasons, they have succeeded in keeping these ‘‘windfall’’ gains out of the generally redistributive income system for Norwegian municipalities. The information that can be extracted from them in the present context is
167
168
The Journal of Human Resources
admittedly limited, however, as they are few and relatively small, and employ only
0.6 percent of the nurses in our study (around 240 persons).
Finally, we have collected (imperfect) information about the cost efficiency of the
municipalities’ nursing sector and their privatization strategies. The efficiency score
measures are collected from Erlandsen et al. (1997) and Edvardsen, Førsund, and
Aas (2000). They refer to two different points in time (1995 and 1997) and are based
on detailed accounts of costs and output in the nursing sector of each municipality.
The scores are computed by means of Data Envelopment Analysis (DEA), and, in
short, a municipality is considered to be fully efficient in the year under consideration if no convex combination of other municipalities exists that produces more care
services along all product dimension, without also using more of at least some inputs.
Our privatization indicator is collected from the so-called KOSTRA database (administered by Statistics Norway), and measures for each municipality in Norway
the fraction of total social service expenditures that were allocated through private
companies in 2003. Although both efficiency scores and privatization indicators
are measured with substantial errors (both because these concepts are difficult to
measure per se, and because they are measured for just a few years, one of them well
outside our observation window), we believe they may provide some information regarding the reform-strategies that were pursued during the 1990s, and hence the pace
of organizational change.
In addition to identifying the behavioral effects of organizational change, we also
wish to examine the possible impact of labor market tightness on the nurses’ employment and absence behavior. Because the nurses in our study typically enjoy very
strong employment protection, we expect discipline effects related to changes in
labor market tightness to be moderate, perhaps with an exception for nurses with
relatively small part time jobs (who might be underemployed). Unfortunately, the
standard local unemployment rates do not represent the employment opportunities
of nurses very well. We have therefore computed monthly county-specific unemployment rates for nurses in Norway particularly for the current project, based on the
whole population of certified nurses in Norway (not only the permanent contract
nurses covered in our analysis). These unemployment rates cover all kinds of unemployment registered at the employment offices.
IV. The Model: The Dynamics of Absence and Welfare
Dependency
In this section, we set up a competing risks hazard rate model aimed
at uncovering the causal effects of organizational change on the propensity to become and remain absent, and to become dependent on social insurance on a more
lasting basis. The model is set up in terms of monthly transition probabilities between
various labor market states, conditioned on vectors of observed and unobserved explanatory variables. Individuals can move among four mutually exclusive states:
(a) Present at work without reliance on social security benefits (this state also
includes absence spells shorter than 16 days, which are paid for by the
employer).
Røed and Fevang
(b) Absent from work with sickness benefits from the NIA.
(c) Nonemployed (or partly employed) with social security benefits (unemployment benefits, rehabilitation benefits, disability benefits, social assistance).
(d) Nonemployed without social security benefits.
In addition, individuals are allowed to move between different workplaces (given
that they are in State a or b). To start with (October 1992), all individuals belong to
one of the first two states; they are either present at work (State a), or absent due to
sickness (State b). Hence, the first part of our model is a binomial probability equation determining the initial state. In the subsequent months, transitions may occur.
Let t=1,2,.,97 be the 97 months for which we can (potentially) observe labor market states, where t=1 is the first month (October 1992). A transition is recorded whenever the state (or workplace) by the end of one month is different from the state (or
workplace) by the end of the previous month. In total, our state space allows for 13
different transition types (12 between the four states described above and one between different workplaces). However, in order to arrive at a tractable and economically interpretable model, we restrict some transitions with a common destination to
be governed by the same causal mechanisms regardless of the current state. In particular, we assume that both observed and unobserved covariates have the same effect on the hazard rates to the two nonemployment States (c and d) regardless of the
current state being presence (State a) or absence (State b). However, sickness absence is allowed to cause duration-dependent shifts in the levels of the hazard rates.
These effects are modeled by means of a timing-of-events approach (Abbring and
Van den Berg 2003a), implying that we do not need exclusion restrictions in order
to identify the causal effects of interest.7 In a similar fashion, we assume that observed and unobserved covariates have the same effect on the hazard rates from nonemployment and back to work presence, regardless of the current state being other
benefits (State c) or self-supported nonemployment (State d). However, the level
of the hazard rate may depend on whether the original exit involved a lasting benefit
or not. We do not model transitions between the two nonemployment States (c and
d); such transitions are only imperfectly observed and disregarded in the analysis.
Note also that we rule out transitions from the two nonemployment states directly
to sickness absence (from State c or d to State b). To the extent that such transitions
occur in the data, we interpret the sickness spell as a continuation of the nonemployment spell until work presence (State a) occurs. Based on these considerations, we
end up with the following distinct modeled transitions [the number of transitions
actually occurring in the data set in brackets]:
Transition 1: From work presence (State a) to long-term sickness absence (State b)
[73,493]
Transition 2: From long-term sickness absence (State b) to work presence (State a)
[61,450]
7. Note that according to Norwegian legislation, sickness absence cannot last longer than 12 months. At
this point some kind of transition must therefore take place.
169
170
The Journal of Human Resources
Figure 3
The modeled transitions (State a). The group ‘‘Not Employed’’ includes some
individuals who have become partly disabled, and have some (part-time)
employment (these are placed in the ‘‘other benefits’’ category).
Note: In addition to the six transitions illustrated in the figure, the initial condition (present or absent)
is also modeled as an endogenous event.
Transition 3: From a job at one workplace (State a or b) to another workplace
[20,683]
Transition 4: From a job (State a or b) to nonemployment benefits (State c) [7,486]
Transition 5: From a job (State a or b) to self-supported nonemployment (State d)
[6,617]
Transition 6: From nonemployment (State c or d) back to a job (State a) [7,576]
The transition pattern is illustrated in Figure 3. Individuals are included in the data
set as long as they do not make a transition to a job outside the workplaces covered
by our longitudinal data (in which case the spells are censored after the recording of
Transition 3). Spells are also exogenously censored for nurses who give birth to a
child, leave the country, or die. As and when these individuals return to the covered
workplaces, their spells are reactivated.
The hazard rates are assumed to depend on six groups of covariates; (i) calendar
time (t); (ii) individual and environmental characteristics (xit); (iii) municipality/
county characteristics (mit); (iv) workplace characteristics (zit); (v) duration of an ongoing sickness absence spell (dit); and (vi) unobserved individual (time-invariant)
characteristics (vi). The explanatory factors are allowed to vary between months
(except unobserved characteristics), but are assumed to be constant within months.
The hazard rates are specified within the framework of a multivariate mixed proportional hazard rate model (MMPH); see Abbring and Van den Berg (2003a; 2003b).
Røed and Fevang
Let uk ðwkit ; vi Þbe the hazard rate governing transitions to state k, given an index
value computed from observed covariates and model parameters wkit, and an unobserved (time-invariant) covariate vki. We specify the hazards in the following
way:
!
uk ðwkit ; vki Þ ¼ exp bk xit + rkt + lkd + ak mit + dk zit + vki ;
|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl}
wkit
ð1Þ
k ¼ 1; 2; :::; 6;
t ¼ 2; 3; :::; 97;
d ¼ 0; 1; :::; 12:
The exponential link-function in Equation 1 is used to ensure nonnegativity of the
hazards uk without constraining the permitted space for the parameters of interestðbk ; rkt ; lkd ; ak ; dk Þ. Note that since all explanatory variables and parameters
are assumed constant within each month, the hazards in Equation 1 can also be interpreted as integrated over these months. The parameters rkt and lkd are the calendar
time and duration effects, represented in the empirical model for transitions 1–6 by
separate dummy variables, one for each calendar month, and one for each possible
sickness absence duration. For transition 6 (from nonemployment and back to work),
more restrictive assumptions are required to estimate the model, given the much
smaller number of observations (the calendar time effects are assumed to be zero
and the duration effects are constrained to be piecewise constant in four-month periods). The initial condition is modeled in terms of a probability function l(.). Let y0i be
an indicator variable taking the value 1 if individual i was absent from work (due to
sickness) by the end of October 1992, and 0 otherwise. We then write the initial condition equation as
0
0
0
ð2Þ probðy0i ¼ 1Þ ¼ l xi1 b0 + m1 a0 + z1 d0 + v0i :
Let ykit 2 Y i be equal to 1 if month t involved a transition to state k for individual i,
and 0 otherwise, where Yi is the complete set of observed transition outcome indicators for individual i. Let K it be the set of feasible transitions for individual i in
month t; K it ¼ f2; 3; 4; 5g for individuals absent from work due to sickness,
K it ¼ f1; 3; 4; 5g for individuals present at work, and K it ¼ f6g for individuals without a job. We can then write the likelihood contribution formed by individual i, conditional on index functions for all relevant k and t ðwi Þand on the six-dimensional
vector of unobserved variables ðvi ¼ ðv0i ; v1i ; v2i ; v3i ; v4i ; v5i ; v6i ÞÞ as
ð3Þ
Li ðwi ; vi Þ ¼ P1 ðwi ; vi Þy0i ½1 P1 ðwi ; vi Þð1y0i Þ
Y Y
ykit 2Y i k2K it
½P2 ðwi ; vi ÞP3 ðwi ; vi Þykit ½1 P2 ðwi ; vi Þ
where
P1 ðwi ; vi Þ ¼ lðw0i ; v0i Þ
is the initial conditions probability Equation 2,
ð1 + ykit Þ
k2K it
;
171
172
The Journal of Human Resources
!
P2 ðwi ; vi Þ ¼ 1 exp + uk ðwkit ; vki Þ
k2K it
is the probability of making any of the available transitions in Kit, and
P3 ðwi ; vi Þ ¼
uk ðwkit ; vki Þ
+ uj ðwkit ; vki Þ
j2K it
is the conditional probability that the transition is of type k.8
Because the likelihood contribution in Equation 3 contains unobserved variables,
it cannot be used directly for estimation purposes. We approximate the unobserved
joint heterogeneity distribution in a nonparametric fashion with the aid of a discrete
distribution. Let Q be the (a priori unknown) number of support points in this distribution and let fvl ; pl g; l ¼ 1; 2; :::Q; be the associated location vectors and probabilities. In terms of observed variables, the likelihood function is then given as
ð4Þ
L¼
N
Y
i¼1
E½Li ðwi ; vi Þ ¼
vi
N
Y
Q
Q
+ pl Li ðwi ; vl Þ; + pl ¼ 1:
i¼1 l¼1
l¼1
To estimate the model, we also have to specify a functional form for the initial
condition probability function l(.) in Equation 2. For computational convenience,
we have chosen to specify it as a complementary log-log function, that is,
lð:Þ ¼ 1 expð expð:ÞÞ.9 Our estimation procedure is to maximize the likelihood
function in Equation 4 with respect to all the model and heterogeneity parameters
repeatedly for alternative values of Q. The nonparametric maximum likelihood estimators (NPMLE) are obtained by starting out with Q = 1, and then expanding
the model with new support points until the model is ‘‘saturated’’ in the sense that
it is no longer possible to increase the likelihood function by adding more points
(Lindsay 1983; Heckman and Singer 1984). At each stage of the estimation process,
we examine the appropriateness of an additional mass-point by means of simulated
annealing (Goffe, Ferrier, and Rogers 1994).10 Monte Carlo evidence presented in
Gaure, Røed, and Zhang (2005) indicates that there is little reason to believe that parameter estimates obtained this way are inconsistent or that they are nonnormally
8. The probability of making a transition to state k between (t-1) and t, given the feasible set of transitions
K (and given that hazard rates are assumed constant within each month), can then be derived as
Z t
Z u
Z t
uk ðwkit ; vki Þ exp +k2K
uk ðwkit ; vki Þds du ¼
uk ðwkit ; vki Þ exp +k2K uk ðwkit ; vki Þ
t1
t1
t1
u ðw ; v Þ
kit ki
k
:
3ðu ðt 1ÞÞ du ¼ 1 exp +k2K uk ðwkit ; vki Þ
+k uk2K ðwkit ; vki Þ
9. This simplifies the maximisation of the likelihood function, since it implies that the initial condition
obtains the same mathematical structure as the hazard-based elements of the likelihood. The complementary
log-log function has the unfortunate property that it is not symmetric. The results may therefore depend on
which outcome is defined as the base state. However, for the small event probabilities that are predominant
in our initial condition, the complementary log-log function is approximately equal to the logit-function.
10. Localisation of the nonparametric maximum-likelihood estimators is an enormous computational task,
and it could not have been done without extensive support from the High-Performance Computing Centre of
the University of Oslo. Our optimisation algorithm is described in detail in Gaure, Røed and Zhang (2005).
Røed and Fevang
distributed. They also indicate that the standard errors conditional on the optimal
number of support points are also valid for NPMLE.
A complete list of observed explanatory variables used in each of the seven equations (the initial condition and the six hazard rates) is provided in Table 3.11 Note that
changes in the level of workplace employment are allowed to have three different
types of effects on the various hazard rates in our model. First, current (ongoing)
changes in the employment level potentially affects all hazard rates; to sickness
absence (and back again), to other jobs, and to nonemployment (with or without
benefits).12 Second, some of these effects may persist after the changes are implemented. We allow these ‘‘lag-effects’’ to last for one year after implementation,
and also to differ between ‘‘survivors’’ (those who stay at the same workplace after
the changes) and ‘‘movers’’ (those who change to a new workplace, for example, as a
result of downsizing). Third, we may expect that the conditions (in terms of employment changes) under which some nurses leave employment (that is, make a transition
of Type 4 or 5) have a causal impact on their propensity to return to work. Note that
effects of downsizing on sickness absence behavior do not necessarily reflect effects
on individuals’ health; see Section II. They also may reflect effects on the threshold
at which a given health status is deemed to justify sickness absence and on the extent
to which sickness is untruthfully reported to conceal other reasons for absence.13 It is
not obvious whether s