J Popul Econ (2004) 17:387–408

DOI 10.1007/s00148-003-0165-y

Assessing the effects of an early retirement program

Espen Bratberg1, Tor Helge Holmås1, Øystein Thøgersen2
1
Department of Economics, University of Bergen, Fosswinckelsgate 6, N-5007 Bergen,
Norway (e-mail: {Espen.Bratberg; Tor.Holmas}@econ.uib.no)
2
Department. of Economics, Norwegian School of Economics and Business Administration,
Helleveien 30, N-5045 Bergen (e-mail: Oystein.Thogersen@nhh.no)

Received: 25 April 2002/Accepted: 15 May 2003

Abstract. We investigate induced retirement effects of the Norwegian early

retirement program ‘‘AFP’’ and emphasize effects caused by relocations of
some individuals from disability pension and unemployment to AFP. Theo-
retical considerations predict that AFP unambiguously induces more early
retirement. Analyzing Norwegian register data 1994–96 with parametric and
non-parametric methods, we demonstrate that i) economic incentives influ-
ence the retirement decision, ii) there is a significant net induced retirement
effect, iii) by a conservative judgment, at least 50% of the AFP retirees would
have stayed in the labor force without the scheme.

JEL classification: J26, H55, J22

Key words: Induced retirement, social security, matching estimators

1. Introduction

Since the early 1970s we have witnessed a growing availability of various

early retirement options in almost all OECD economies.1 In many countries

All correspondence to Espen Bratberg. We are indebted to the referees for detailed remarks,
which significantly improved the paper. Financial support from the Norwegian Research Council
and the Ministry of Health and Social Affairs is greatly appreciated. We are grateful for valuable
comments from Erik Hernæs and Astrid Grasdal, seminar participants at the Norwegian School
of Management in Oslo, the Institute for International Economic Studies in Stockholm, and the
University of Linz, the 2000 Conference of the European Society for Population Economics in
Bonn, and the German-Norwegian Seminar on Social Insurance in Berlin, 2000. Bratberg would
also like to thank the Humboldt University for its hospitality during a stay in the winter of 2003.
Responsible editor: Christoph M. Schmidt.
388 E. Bratberg et al.

these options have included a liberalization of the entitlement conditions for

older workers’ access to disability and unemployment programs (OECD
1998a). Moreover, many economies have introduced voluntary early
retirement programs implemented as an extension of the ordinary public
pension scheme or negotiated between the unions, the firms and potentially
the government as a third part. According to OECD studies (OECD 1998a,
1998b) most of these schemes give strong economic incentives to withdraw
from the labor force before the standard pension age.
Focusing on the latter type of voluntary programs, we note that they
have been rationalized by a postulated need for a flexible retirement age
interpreted as both the option to retire before the standard retirement age
and the opportunity to remain in the labor force for some additional years.
Data for the OECD economies suggest that this flexibility works only one
way, however. As illustrated in the OECD studies cited above, the average
retirement age has declined steadily over the last couple of decades. Con-
sequently, we may suspect that there are significant induced retirement
effects associated with most voluntary early retirement schemes. Assessing
these effects, we must disentangle the net ‘‘real’’ induced retirement effects
from the effects caused by possible relocations of some individuals from
other transfer programs (notably disability pensions and unemployment
benefits). This crucial issue is particularly highlighted in the present paper.
The observed escalation of early retirement increases the well-known and
potentially severe problems related to ageing populations in all OECD
economies. Early retirement implies reductions in the labor force and con-
sequently a lower potential GDP. In addition the associated increase in the
ratio between the number of social security receivers and labor market par-
ticipants threatens the financial viability of these economies’ social security
systems, which essentially are financed on a pay-as-you-go basis.
This paper considers the effects of the fairly generous Norwegian early
retirement program ‘‘AFP’’, which was introduced in 1989. Because AFP-
eligibility is firm-specific, it may be argued that the introduction constitutes
a natural experiment, with workers in non-affiliated firms as a control
group. This property hinges on the presumption that there are not sys-
tematic differences in unobserved characteristics of workers in firms that
are affiliated to the AFP-scheme and workers in firms that are not, and
that workers do not change firms to obtain the AFP-option. As time
passes and workers and firms have time to adapt, the property is gradually
lost. However, we shall argue below that in the period under study in this
paper (1994 – 1996), the AFP program did have the properties of a quasi-
natural experiment. This alleviates the general identification problem for
causal analysis caused by the possibility that underlying variables, which
determine variations in economic incentives, also may have an independent
direct effect on retirement decisions, see Krueger and Pischke (1992:
p. 415).
Based on theoretical considerations we predict that the AFP–program’s
favorable economic incentives should unambiguously induce more early
retirement. Because the net replacement rate of the AFP program is a
declining function of income, it also follows that the substitution effect to-
wards early retirement is stronger for low-income individuals than high-in-
come individuals. Using a large longitudinal sample of old individuals
employed in the private sector, we test these predictions. We also make a
Early retirement 389

point of assessing to what degree the new retirement scheme only substitutes
for already existing ones like disability pension.
This paper adds to the large literature on retirement behavior, see
Lumsdaine and Mitchell (1999) for a broad survey. It is particularly closely
related to other analyses of retirement behavior, which shed light on the
degree of substitution between voluntary programs and other schemes (dis-
ability, unemployment) and/or attempt to utilize data from developments
interpreted as natural experiments.
Using Danish data, Pedersen and Smith (1996) analyze transitions from
work to respectively disability and a voluntary early retirement scheme
(‘‘Etterlønn’’). The latter scheme has fairly strong similarities with the Nor-
wegian AFP program. Pedersen and Smith show that there are different sets
of factors that explain exits to these retirement pathways. We note in par-
ticular that a higher replacement rate stimulates exits through the voluntary
scheme but not through disability. A Finnish study by Lilja (1996) also
confirms that different sets of factors explain transitions through different exit
routes. Lilja’s results indicate that high-income groups retire later than low-
income groups.
A paper utilizing a natural experiment is Krueger and Pischke (1992).
They consider a reform of the US social security system in 1977, which lead to
an unexpected significant drop in benefits to generations born after 1916
(compared to generations born just before this year). In contrast to several
earlier studies of the US social security system, their analysis indicates an
insignificant relationship between social security wealth and labor supply.
Dahl et al. (2000) consider retirement behavior in Norway. They analyze
transitions from work to either disability pensions or unemployment benefits
and identify various push and pull factors that explain transitions to these
destinations. Their analysis indicates that disability and unemployment are
not exchangeable pathways. Moreover, economic incentives seem to be of
some importance for exits to unemployment and disability. However, Dahl
et al. do not have data on AFP early retirement.
Two recent papers, Hernæs et al. (2000) and Røed and Haugen (2003),
offer analyses of the Norwegian AFP scheme. Hernæs et al. (2000) estimate a
competing risk model for transitions from full time work to full AFP
retirement or to the combination of partial AFP retirement and part-time
work. They find that economic incentives influence the early retirement
decision significantly, and by means of simulations they demonstrate that a
rather generous ‘‘bonus’’ is necessary to make individuals postpone retire-
ment for one year. However, they do not deal with the potential interaction
between the AFP scheme and the other early retirement pathways. As they
only consider individuals who are eligible for AFP, the natural experiment
property of the reform is not addressed. Røed and Haugen (2003) do focus on
the natural experiment property of the reform and find that the introduction
of subsidized early retirement has reduced the employment rate among the
elderly sharply. The present analysis differs from Røed and Haugen by
focusing on the private sector – for reasons discussed later – and in the data
available for identifying AFP-firms.
The next section presents the institutional setting, derives the expected
behavioral effects of the AFP program by means of a simple theoretical model
and takes a closer look at important aspects of the natural experiment ap-
proach. Section 3 outlines the estimation strategy, while Sect. 4 describes our
390 E. Bratberg et al.

data set. Our results are presented and interpreted in Sect. 5. A key finding is
that by a conservative judgment, at least 50% of AFP retirement is induced
by the scheme. Section 6 offers some final remarks.

2. Institutional setting and the AFP scheme

2.1. Institutional setting

The standard retirement age in Norway is 67, when old age pension benefits
are generally available. The AFP early retirement program was introduced in
1989 as the result of negotiations between employers, unions and the
government. At the outset the entitlement age was set to 66 but has gradually
been lowered to 65 in 1990, to 64 in 1993, to 63 in 1997 and finally to 62 since
1998. The AFP program covers the public sector and almost half of the
employees in the private sector. This reflects that only private firms that
participate in the central tariff agreements – mainly those organized by the
employers’ organization NHO (Confederation of Norwegian Business and
Industry) – can offer AFP to their employees.
Prior to the introduction of AFP, the available early retirement routes
were disability benefits (subject to a medical test), prolonged unemployment
benefits or sickness benefits. These routes were in principle not voluntary, but
they have still been available for significant segments of older workers
through unofficial practice - particularly during recessions (see Dahl et al.
2000). During the deep recession in the early 1990s, a liberalization of the
official disability requirements took place as well.2 Consequently, an argu-
ment put forward in favor of the AFP-reform was that some workers would
not stay in the labor force until 67 anyway, and the main effect of the new
program would be to offer them a more ‘‘dignified’’ exit from the workforce.
For workers in the public sector or in AFP-affiliated private firms, AFP
early retirement requires only that some rather weak conditions for labor
market participation are satisfied. These requirements are essentially i) 10
years after the age of 50 with income at least equal to 1G (‘‘G’’ is the basic
counting unit in the social security system, regulated annually and 37033
NOK in 1993, the base year for our calculations of potential pensions), ii) 10
years with income at least at 2G since 1967, iii) annualized income at least
equal to 1G in the calendar year of AFP retirement and in the year before and
iv) at least three years of employment in the present firm. In March 2003,
1 EURO » 7.8 NOK.
The calculation of the AFP early retirement benefits is closely related to
the calculation of regular public pension benefits, which the individual re-
ceives after the age of 67.3 Thus, in order to assess the economic incentives of
the AFP program, we must first consider the regular public pension system in
Norway. This is a mandatory defined-benefit system, which includes an
earnings-based supplementary benefit in addition to a fixed minimum pension
benefit. The latter accrues to all individuals regardless of any participation in
the labor market at all. In 1993 the minimum pension benefit amounted to
nearly 60000 NOK. Essentially, the minimum pension is tax-free, while the
supplementary pension is subject to taxation.4
The magnitude of the supplementary pension depends on the earnings
level of the 20 ‘‘best’’ years and the length of the working career. The latter
Early retirement 391

counts in the sense that the supplementary pension is reduced proportionally

if the individual has less than 40 years of earnings above the fairly low
minimum level of 1G or 37033 NOK (in 1993). Claims to supplementary
pension are accumulated in terms of so-called ‘‘earning points’’, which in turn
is translated into money amounts. We note that only gross labor income
above a certain annual level (approximately 91000 NOK in 1993) leads to
accumulation of claims to supplementary pension. Above this level there is a
positive and piecewise linear relationship (with a gradually flatter slope) be-
tween gross income and pensions. There is a ceiling in the sense that a gross
income of approximately 445000 NOK yields a maximum total public pen-
sion benefit equal to approximately 176000 NOK in 1993.
The AFP early retirement benefit is calculated as the public pension
benefit the individual would have received at age 67 – assuming hypothetically
that he had continued in the work force until that age. This means that a
projected labor income stream until age 67 (and the associated earning points)
is taken into account in the calculation of the supplementary pension.5 In
addition the individual also receives an ‘‘AFP-subsidy’’ during the early
retirement years. In the private sector this subsidy is tax-free and amounts to
an annual 11400 NOK in 1993. When an AFP pensioner reaches the age of
67, he will receive an ordinary public pension, which is still calculated as if he
had continued in the labor force until age 67. It follows that the government
contributes to the AFP scheme in the private sector both indirectly through
‘‘free’’ earning points for the regular supplementary pension and directly
through the financing of 40% of the benefits to individuals aged 64, 65 and
66. The remaining expenditures are financed by the employers.
Considering full AFP early retirement in the private sector in 1993, the net
replacement rate exceeds 100% for a low annual gross income of approxi-
mately 100000 NOK. A higher income level yields a lower net replacement
rate, which for example amounts to 65% for a medium annual income of
250000 NOK and 50% for a fairly high annual income of 400000 NOK. We
note that the combination of part time employment and part time AFP leads
to somewhat higher net replacement rates.6

2.2. Economic incentives

The main incentive effects of the AFP scheme can be illustrated by means of a
simple model. We imagine for the rest of this section that a representative
individual has a fixed life span normalized to unity. This individual chooses
his retirement age a (0<a<1) in the sense that he participates inelastically in
the labor force in the first a part of the life span, while the last l = 1–a part is
spent in retirement. There is a standard retirement age a
. A voluntary early
retirement scheme (i.e. AFP) is available after age a
,
a  a
. We disregard the
rare possibility of a > a
and assume that both the real interest rate and the
individual’s time preference rate are zero.
The optimal choice of a is determined by the maximization of the utility
function
U ¼ U ðcðaÞ; lðaÞÞ ð1Þ
where c is consumption. Utility is strictly increasing and concave in both its
arguments. The maximization is subject to the constraint a 
a and
392 E. Bratberg et al.

cðaÞ ¼ aW net þ pðaÞ ð2Þ

net
Here w is an exogenous net wage per time unit and pðaÞ is a pension benefit.
Following essentially Fehr et al. (2003), the pension formula is given by


pðaÞ ¼ ð1  aÞ A þ f ða; wgross Þ þ wða
 aÞ; ð3Þ
where wgross is the gross wage rate. The first term on the RHS of (3) captures
that the individual receives a flat benefit A (A>0) and a net of tax
supplementary benefit f ða; wgross Þ during retirement, f ða; wgross Þ  0, @f @a  0;
@f
@wgross  0: The second term on the RHS of (2) reflects that the individual faces
an additional early retirement subsidy (w>0) during the early retirement
period. Clearly, the constraint a 
a implies that we disregard the possibility
that individuals may retire before
a without receiving any pension benefit
before a
.
The pension formula (3) captures the important incentives for Norwegian
individuals eligible for AFP early retirement. As noted above, the early
retirement benefit is calculated as the pension benefit the individual would
have received at age 67 plus an additional early retirement subsidy. The latter
subsidy corresponds to w > 0. The fact that the supplementary pension level
is calculated as if the individual had continued in the labor force until the
standard retirement age corresponds to @f =@a ¼ 0; provided that the indi-
vidual has reached the AFP entitlement age of
a.
Assuming first that the constraint a 
a does not bind, it follows from (1),
(2) and (3) that the first-order condition for the optimal choice of a is
@U @U @c

b ðaÞ ¼ ;b  ¼ wnet  A þ f ða; wgross Þ  w: ð4Þ
@c @l @a
Thus, the marginal utility of more consumption, caused by a longer career in
the labor force, should equal the marginal utility of a longer retirement period.
@c
We interpret b  @a as the ‘‘price of a prolonged retirement period’’ because
this derivative expresses the marginal price of a longer retirement period in
terms of forgone consumption. The magnitude of b is constant during the
interval from
a to a
due to the property @f@a ¼ 0 (and will indeed increase after


a because the early retirement subsidy w no longer applies, see (3)).
We observe from (4) that a partial increase in A, w, or the magnitude of
f ða; wgross Þ will reduce b. Moreover, it is straightforward to verify that this
will lead to both substitution and income effects towards a lower retirement
age. Because the net replacement rate is a declining function of income in the
Norwegian system, the b values will be low (and indeed close to zero) for low-
income groups, i.e., these groups will face the strongest substitution effects
towards early retirement.
By means of our theoretical set-up, the introduction of the AFP program
(or a reduction in the AFP entitlement age) can be interpreted as a reduction
in
a. Accordingly, an individual has an incentive to retire earlier if the
restriction a 
a binds at the outset, implying that
@U @U
b ðaÞ < ðaÞ: ð5Þ
@c @l
We can imagine an individual, who initially - before the introduction of AFP -
could not retire before the standard age a
(without accepting zero income in
Early retirement 393

the meantime). Then the introduction of AFP implies that a
is lowered below
a
and the price of a prolonged retirement period is no longer wnet but b as
calculated in (4). This implies a negative substitution effect on a, which is not
offset by any income effect. Thus, theory predicts that the AFP scheme
unambiguously induces more early retirement. We also expect that the
tendency to choose AFP declines with the income level, reflecting the inverse
relationship between the net replacement rate and the income level in the
Norwegian system.
Given the favorable economic incentives and the fairly weak individual
eligibility conditions, it is not very surprising that the AFP program has
gradually gained increased popularity in response to the reductions in the
entitlement age and a growing knowledge of this fairly new scheme among the
workers. According to a Norwegian government white paper (NOU 1998:19),
AFP has been the most utilized exit path from the labor force for individuals
aged 64–66 years since 1993. Moreover, the AFP program accounts for
approximately the same amount of exits as the disability program when we
consider the larger age group 60–66 years.

2.3. AFP as a natural experiment

The basic assumption underlying our natural experiment approach is that

workers in AFP and non-AFP firms do not differ systematically in their
unobserved characteristics. Because we cannot rule out that people with
different preferences self-select into the private and public sectors, we
consequently choose to focus exclusively on individuals employed in the
private sector. By including the public sector, a majority of the AFP-eligibles
would be public employees. This would increase the possibility that AFP
entitlement is correlated with unobserved characteristics.
Even if the public sector is excluded, we must still disregard the possibility
that workers in the private sector switched from non-AFP to AFP firms in
order to obtain the AFP option. Recalling that AFP early retirement requires
three years of employment in the present firm, this potential problem is re-
duced by the facts that i) our sample includes only individuals who reached
the age of 64 during the first half of 1994 (see more details below) and ii) the
AFP scheme was introduced in 1989. As documented by Salvanes and Førre
(2003), the job market mobility of elderly workers is extremely low in gen-
eral.7 Thus, it seems hard to imagine that a significant share of the workers in
our sample switched job as a part of an ‘‘early retirement strategy’’ during the
fairly short time span from 1989 to the first half of 1991 - a period which was
even characterized by a recession. This view is also supported by Røed and
Haugen (2003). They investigated whether workers who became eligible for
AFP in 1997 switched jobs 3–4 years prior to eligibility, and found that such
job shifts were extremely rare.
Additional potentially crucial aspects are related to whether firms in effect
choose AFP affiliation or not due to strategic considerations, and to whether
firms’ affiliation to AFP is related to the unobservables of the individual. The
fact that firms’ AFP affiliation is a consequence of their more fundamental
decision to be affiliated with the employers’ organization NHO, contributes to
reducing the probability of systematic changes in affiliation caused by early
retirement concerns. Looking at the development in the total number of firms
394 E. Bratberg et al.

affiliated with the NHO (according to their annual reports), we neither ob-
serve any indication of widespread changes in affiliation due to the intro-
duction or the various extensions of the AFP scheme. In fact the number of
firms affiliated with the NHO seems to be very closely related to the general
business cycle developments.
Because firms that are not part of the centralized tariff negotiations (i.e.,
the non-AFP firms) are typically small and also unevenly distributed across
sectors, it still cannot be ruled out that workers in AFP and non-AFP firms
differ in characteristics. Our remedy is to condition on the observable char-
acteristics that are available. Below we also present some evidence from our
data source, which we think is in favor of the identifying assumption. Before
that, the next section outlines our approach to estimating the effect of
introducing this early retirement program.

3. Empirical approach

3.1. The evaluation problem

This section describes the non-parametric and parametric estimators that will
be applied in the analysis. We consider several discrete labor market
outcomes. Denote the set of potential outcomes (more precise definitions
follow in the next section) for an individual with AFP entitlement J1 =
{work, AFP, disabled, sick, other}, and the according set for an individual
without entitlement J0 = {work, disabled, sick, other}. We ignore time-
indexing and define a set of dummy variables Yij1 ¼ 1 if an entitled individual
i’s outcome is j ˛ J1, 0 otherwise. Accordingly, Yij0 ¼ 1 if the outcome of
an individual without entitlement is j ˛ J0, 0 otherwise. Thus EðYijs Þ ¼
PrðYijs ¼ 1Þ  Pijs ; s ¼ 0; 1: The interesting question is how the expected
outcome is affected by AFP eligibility, apart from the trivial fact that the
probability of AFP is zero without entitlement. However, this treatment
effect (or rather, eligibility effect) is intrinsically unobservable because it
considers a difference between two potential outcomes, and any individual
can only be either eligible or non-eligible at any point in time.
Define a new dummy variable Ti = 1 if individual i is eligible (or ‘‘trea-
ted’’), 0 otherwise. Estimation of treatment effects typically focuses either on
the average treatment effect or the effect of treatment on the treated. The
average treatment effect is DAT ¼ EðYij1 Þ  EðYij0 Þ. This is the treatment effect
for a random individual in the population. More often, the interest is on the
effect of treatment on the treated, DTT ¼ EðYij1 jTi ¼ 1Þ  EðYij0 jTi ¼ 1Þ. In our
case, we are considering the effect of introducing AFP on those eligible,
corresponding to treatment on the treated. In estimation, the problem is to
find a proxy for the counterfactual EðYij0 jTi ¼ 1Þ, i.e., the expected outcome of
an entitled individual if she/he were not eligible.
In a randomized experiment, Ti would be randomly assigned, and
EðYij0 jTi ¼ 1Þ could be proxied by EðYij0 jTi ¼ 0Þ, i.e. the behavior of the
experimental control group could replace the behavior of the treated group in
absence of treatment. It is hard to imagine a randomized experiment in the
AFP context; in any case such data are not available (see also Heckman and
Smith 1995, for a general discussion on problems related to social experi-
ments). We have argued above, however, that the introduction of AFP may
Early retirement 395

be considered as a natural experiment, though not a perfect one. The iden-

tifying assumption is that there is no selection on unobservables into the
group of AFP-eligibles, or Yij0 is distributed independently of Ti. This
assumption is not testable, but we shall present some evidence in its favor
later. If it holds true, the non-eligibles constitute a comparison group. Based
on this assumption, we present non-parametric and parametric estimators of
the effect of AFP-entitlement.

3.2. Non-parametric estimators

Assume that Yij0 is distributed independently of Ti. In that case, the difference
^ D ¼ EðY 1 jTi ¼ 1Þ  EðY 0 jTi ¼ 0Þ
D ð6Þ
TT ij ij

is an unbiased estimator of the AFP-effect. This is the standard estimator in

randomized experiments, which we label ‘‘unweighted difference’’ in the next
section.
Assume (more weakly) that conditional on a vector of observable char-
acteristics Xi, Yij0 is independent of Ti. This is the conditional independence (CI)
assumption. When CI holds, an estimator may be constructed by matching
the treated and non-treated on X and comparing matched individuals in the
two groups. When the dimensionality of X is high, or it contains many
continuous variables, matching may be facilitated by conditioning on the
propensity score. The propensity score, p(Xi) = Pr(Ti = 1|Xi), is the prob-
ability of treatment conditional on X. Rosenbaum and Rubin (1983) show
that if CI holds, Yij0 is distributed independently of Ti conditional on p(Xi).
The average effect of treatment on the treated may then be estimated as
^ M ¼ E½EfY 1 jTi ¼ 1; pðXi Þg  EfY 0 jTi ¼ 0; pðXi ÞgjTi ¼ 1;
D ð7Þ
TT ij ij

where the outer expectation is over the distribution of p(Xi) in the treated
population.
In practice, p(Xi) must be estimated, e.g. by logit or probit. It is important
that the distribution of X conditional on the estimated propensity score is
balanced in the sense that the covariates are distributed as similarly as pos-
sible in the treatment and comparison groups. We follow an algorithm for
estimating the propensity score described by Dehejia and Wahba (2002):
based on an estimated propensity score ^pi , the groups are divided into
intervals such that the averages of the score are not significantly different
between the groups. Within each interval, the means of X in the two groups
are tested for equality. If the tests fail, the estimated score may be improved
by adding interactions and/or higher order terms of the variables in X.
Matching estimators are computed by comparing each member of the
treatment group to the comparison group members using some weighting
scheme based on the propensity score. For an extensive discussion of matching,
see Heckman et al. (1998). In this paper we present the following estimators. i)
Nearest neighbor matching (NN): each treated individual is compared to the
comparison with the closest propensity score. Usually that is done with
replacement, so each comparison group member may be used repeatedly. ii)
Radius matching: each treated individual i is compared to the average of the
comparisons within a radius of ^ pi . iii) Kernel matching: each treated individual i
396 E. Bratberg et al.

is compared to a kernel weighted average of the comparisons, with weights

pi . iv) Interval matching: the sample is divided into
declining by the distance to ^
intervals based on the estimated propensity score. The estimator is the average
difference between the treatment and comparison groups into intervals,
weighted by the fraction of the treatment group that falls into each interval.

3.3. Parametric estimator

We supplement the non-parametric estimators with a parametric estimator,

where we assume that the outcome of individual i may be specified by a
parametric model. More specifically, we assume that
Pij0 ¼ f ðc0 Wij þ g0j 1 Zi Þ; Pij1 ¼ f ðc1 Wij þ g1j 1 Zi Þ; ð8Þ
where f(.) is a probability function, Wij is the income associated with state j,
Zi are other variables characterizing individual i, c is a parameter, and gj is a
state specific parameter vector. Hence it is assumed that the difference in
behavior between those entitled and those not is picked up by different
coefficients. If there is no selection on unobservables into the AFP- and non-
AFP-groups, the effect of AFP may be estimated by first estimating ðc1 ; g1ij Þ
and ðc0 ; g0ij Þ for the AFP and non-AFP-groups (where AFP is not among the
potential outcomes) respectively, and then computing
^ Par ¼ f ð^c1 W j þ ^
D g11 ZjTi ¼ 1Þ  f ð^c0 W j þ ^g0j 1 ZjTi ¼ 1Þ ð9Þ
TT ;j

for j „ AFP. We do this using the multinomial logit model (MNL) to

represent f(.). The MNL may be interpreted as a random utility model with
extreme value distributed error terms. In this context, however, we rather
consider the MNL as a convenient parametric approximation to the
probabilities.8A potential advantage of this approach is that it enables us
to model the effect of economic incentives (but at the cost of introducing
untestable parametric assumptions).

4. Data

4.1. Data source and sample

Our data are collected from the ‘‘MOTIPE’’ database, which covers the
period 1992–1996. This database, which is administered by the Norwegian
National Insurance Administration (NIA), contains individual level infor-
mation on socio-economic background, labor market participation and social
insurance status of the population of individuals aged 62 to 70. It is based
solely on administrative records, supplied by the NIA, Statistics Norway, and
the Directorate of Labor. Thus no survey information is included. MOTIPE
contains AFP payments for receivers, but not on eligibility for individuals
who do not receive AFP benefits. Therefore, for the purpose of this study, our
sample has been merged with a file containing information on which firms
were affiliated with the AFP scheme as of December 1995.9
Our sample is defined as follows.
We consider individuals in the private sector who fulfilled the eligibility
conditions (see above) and reached the age of 64 between January 1 and July
Early retirement 397

31, 1994 (born January 1 – July 31, 1930). The AFP eligibility age was re-
duced from 65 to 64 in October 1993. Thus affiliated individuals in our sample
are eligible from their 64th birthday, and may be followed until the age of 66½
(December 31, 1996 for an individual born July 31, 1930).
Individuals who received disability or unemployment benefits as of Jan-
uary 1, 1994, are excluded (i.e., all individuals in the sample were employed at
that time).
This leaves us with a sample of 3505 individuals, where 1939 individuals
are entitled to AFP, and 1566 are not. At age 66½ they are classified
according to the following labor force states: i) Still in work, ii) AFP pen-
sioner, iii) Disabled, iv) Long term sick leave or v) ‘‘Other’’. i) is based on
records from the Directorate of Labor (the employee register10), and ii) – iv)
on NIA records. We treat disability and long term sickness as different states
because the associated benefits differ (sick leave: full replacement, disability
about 67% replacement). As noted in Sect. 2.1, extended unemployment is a
possible exit route, but at 66½ this outcome constitutes a quite small group
which is placed in the residual category v).11
In the classification of individuals in destination states, we give AFP
priority over disability pension and long term sick leave, in the sense that
individuals with disability and/or sick leave spells before transition to AFP
are classified as AFP early pensioners. Similarly, disability pension is given
priority over long-term sick leave. Individuals who leave the employees reg-
ister and have no transition to ‘‘AFP’’, ‘‘Disabled’’ or ‘‘Long term sick leave’’
are classified as ‘‘Other’’. We note that the ‘‘Other’’ group includes individ-
uals who are retired through private or firm-provided retirement schemes (on
which we have no information), self-employed or out of the labor force for
other reasons.
Table 1 provides descriptive statistics by AFP eligibility. For most of
the individual characteristics we do not observe any large differences between
the subsamples. However, the proportion of males is somewhat higher in the
AFP sample. As noted above, the distribution across industries is quite
different for the subsamples.
Table 1 also gives a first look at the observed outcomes at age 66½.
Clearly, the share that remains in the labor force after 2.5 years is much
higher in the non-AFP sample (63.7%) than in the AFP sample (34.9%).12 On
the other hand, a larger number of individuals in the non-AFP sample have
transitions to disability pension. We also notice differences in transitions to
sick leave and the residual group. As discussed in Sect. 3.2, a simple estimate
of the AFP effect may be computed based on these numbers. Before doing
that, however, we consider the credibility of the identifying assumption given
the available data.

4.2. Evidence regarding selection

The assumption that the outcome is unconfounded with eligibility (Ti) is

crucial but untestable. A comparison of how cohorts born around 1923
behaved in 1989 might have presented strong evidence in its favor: these
individuals would have little time to adapt to the new scheme before their 66th
birthday. However, 1989 is outside the range of MOTIPE. In stead, we
present two comparisons that we think provide evidence, albeit admittedly
398 E. Bratberg et al.

Table 1. Descriptive statistics. Means/percentages (standard deviations/frequencies in parenth-

eses)

All With AFP Without

entitlement AFP ent.

Post-tax income 1993 15.5 (7.3) 15.4 (6.2) 15.5 (8.4)

Calculated net AFP income per year 11.1 (2.9) 11.1 (2.7) 11.0 (3.0)
Calculated net disability pension per year 10.4 (2.9) 10.5 (2.8) 10.4 (3.1)
Calculated net sickness benefits per year 14.6 (4.5) 14.7 (4.2) 14.5 (4.9)
Experience (years with income>1G, 25.3 (3.5) 25.6 (3.3) 25.0 (3.8)
NOK 37,033 in 1993)
Days unemployed 1992 1993 18.4 (81.5) 20.0 (80.5) 16.4 (82.7)
Days on sick leave 1992-1993 18.7 (53.0) 20.6 (55.8) 16.5 (49.1)
Male 69.8 (2446) 74.7 (1448) 63.7 (998)
Married 79.3 (2780) 80.4 (1558) 78.0 (1222)
Full-time worker (‡ 30 hours per week) 80.1 (2806) 82.9 (1607) 76.6 (1199)
Part-time worker (20-29 hours per week) 10.1 (355) 9.1 (176) 11.4 (179)
Part-time worker (< 20 hours per week) 9.8 (344) 8.1 (156) 12.0 (188)
Industry
Agriculture/fisheries 2.9 (101) 2.3 (45) 3.6 (56)
Petrol 1.5 (54) 2.2 (42) 0.8 (12)
Manufacturing 32.1 (1125) 48.3 (936) 12.1 (189)
Construction 5.8 (203) 6.8 (132) 4.5 (71)
Retail 27.8 (975) 23.1 (448) 33.7 (527)
Transport 8.9 (312) 7.1 (138) 11.1 (174)
Finance 8.0 (279) 4.2 (81) 12.6 (198)
Service 13.0 (456) 6.0 (117) 21.7 (339)
Region 1 (Eastern Norway) 38.9 (1355) 32.4 (628) 47.1 (737)
Region 2 (Central Norway) 22.1 (775) 25.6 (496) 17.8 (279)
Region 3 (South/Western Norway) 31.5 (1104) 33.3 (645) 29.3 (459)
Region 4 (Northern Norway) 7.5 (261) 8.8 (170) 5.8 (91)
Outcomes at age 66½ (in 1996)
Still in work 47.8 (1674) 34.9 (677) 63.7 (997)
AFP 27.7 (969) 50.0 (969) –
Disability 9.3 (326) 5.7 (111) 13.7 (215)
Long term sick leave 6.7 (234) 5.3 (103) 8.4 (131)
Other 8.6 (302) 4.1 (79) 14.2 (223)
N 3505 1939 1566

Sample: Individuals born Jan 1 – July 31, 1930.

Measured in 1993 unless otherwise noted.
All incomes in 10,000 1993 NOK.
Calculated pensions/benefits based on 1993 earnings/pension points + the relevant tax, benefit
and pension rules.

less conclusive. First, a comparison of 62-year-olds in 1994 (born in 1932),

based on individuals in the MOTIPE data working in the same firms and
selected in the same way as the main sample, except for the birth dates.
Second, a comparison of individuals who turned 64 in the first three months
of 1993 (born January 1 – March 30, 1929) to those who turned 64 in the first
three months of 1994 (a part of the main sample).
Table 2 shows outcomes at age 63 in AFP- and non-AFP firms for
workers who were 62 in 1994 and thus not (yet) eligible for AFP. If workers in
AFP- and non-AFP firms are systematically selected by unobserved charac-
teristics that affect the labor market outcomes at age 64, one might expect
Early retirement 399

Table 2. Labor market outcomes by AFP affiliation, 63-years-olds

All Working in Working in

AFP firms non-AFP firms

Still in work 1443 (83.0) 999 (82.1) 444 (85.2)

Disabled 97 (5.6) 69 (5.7) 28 (5.4)
Long term sick leave 109 (6.3) 81 (6.7) 28 (5.4)
Other 89 (5.1) 68 (5.6) 21 (4.0)
N 1738 1217 521

Percentages in parentheses.
The table shows labor force status at age 63 for workers born Jan 1 – July 31, 1932 (aged 62 in
1994).

Table 3. Labor market outcomes by AFP affiliation, 64-year-olds in 1993 and 1994

64 years 1. Jan–1. Mar 1993 64 years 1. Jan–1. Mar 1994

Working in Working in Working in Working in

AFP-firms non-AFP firms AFP firms non-AFP firms

Still in work 399 (82.6) 322 (83.9) 664 (64.7) 699 (87.7)
AFP – – 267 (26.0) –
Disabled 15 (3.1) 13 (3.4) 25 (2.4) 40 (4.9)
Long term sick leave 23 (4.8) 11 (2.9) 34 (3.3) 14 (1.7)
Other 46 (9.5) 38 (9.9) 36 (3.6) 60 (7.4)
N 483 384 1026 811

Percentages in parentheses.
Left panel: Labor force status at age 64¼ for workers born Jan 1 – March 30, 1929 (AFP
entitlement age = 65).
Right panel: Labor force status at age 64¼ for workers born Jan 1 – March 30, 1930 (AFP
entitlement age = 64).

that these characteristics affect outcomes for workers who are two years
younger. There are 3% fewer individuals who are working in the AFP firms,
but still the main impression is that the groups are very similar. Against this it
may be argued that with the prospect of eligibility at 64 a 62-year-old may still
change behavior in the years in between. E.g., if there is social stigma asso-
ciated with disability benefits, or if loss of the AFP subsidy is an issue, fewer
workers in AFP firms might want to apply for disability benefits.
Table 3 compares workers in 1993 to workers of the same age in the same
firms in 1994. In October 1993, the AFP age limit was changed from 65 to 64.
We tabulate labor market outcomes for workers who turned 64 January 1 –
March 31 1993, three months after their respective birthdays (when they were
still not eligible), and compare to workers who turned 64 January 1 – March
31 1994. In this sample the percentage still in work in non-AFP firms in 1994
is almost identical to the according percentage for workers in AFP-firms in
1993. Contrary to the previous table, we now compare individuals of the same
age as in the main sample (but with a shorter follow-up). Still, similar
arguments may be put forth against this comparison as for Table 2. However,
we also find a significant difference in the employment percentage in AFP-
firms for 64-year-olds from 1993 to 1994, and 64-year-olds in AFP firms in
1993 are more similar in this respect to workers of the same age in non-AFP
400 E. Bratberg et al.

firms than in AFP firms in 1994. Due to small sample size a comparison of the
disability fraction in the two groups of workers in AFP firms is difficult, but
there is a small reduction. In our view, Table 3 gives some evidence against
endogenous selection of AFP firms. Workers in AFP firms who ‘‘missed’’ the
expansion of the age limit in 1994 actually could constitute an alternative
comparison group, but we do not have a large enough sample to explore that
possibility.

5. Results

5.1. Non-parametric estimates

Table 4 first repeats the observed outcomes in the AFP and non-AFP-group
for convenience, and then reports non-parametric estimates as described in
Sect. 3.2. Starting with the unweighted difference estimator in column 3, it
indicates a significant reduction in employment of 28.8 percentage points due
to introduction of the AFP scheme. There is also substitution from the other
informal early retirement paths, adding up to 21.2 percentage points
(abstracting from the confidence intervals). Thus, according to these
estimates, more than half the AFP pensioners are ‘‘fresh’’ retirees who would
be working in absence of the AFP option.
The estimates in column 3 are unbiased only if entitlement and outcome
are independent variables. Columns 4 –7 improve the estimates by condi-
tioning on observable characteristics, using the propensity score as described
in the previous section.13 The propensity score was estimated by logistic
regression, and the applied model (reported in the appendix) satisfied tests for
balancing. As discussed in Sect. 4, the matching estimators differ in how
matches are selected and weighted, and this is reflected in the results. The
nearest neighbor (NN) estimates stand out as showing the largest effect on the
probability of working, with an estimated 32.2 percentage points reduction.
Estimates for the other transitions are more similar, except for the ‘‘other’’-
state, where the estimates vary from –6.5 for the nearest neighbor estimator to
–8.8 for the radius estimator. Notably, all the matching estimators yield larger
effects on work and smaller effects on the ‘‘other’’-state than the unweighted
estimator. It therefore seems appropriate to infer that the unweighted esti-
mator under- and overestimates, respectively, the effects with respect to those
states. What regards the different matching estimators, one cannot say a
priori which one is better. The NN estimator uses the best possible match for
each observation, but that may still be a poor one, and ‘‘good’’ and ‘‘bad’’
matches are weighted equally. The radius estimator may disregard very bad
matches, even though that is not the case here, as all units in the AFP-sample
have a match. The kernel estimator in general uses all the information in the
control group, but in our case the number of used controls is not affected. It
seems that a reasonable attitude to the reported results is to combine the
confidence intervals to form conservative assessments of the AFP effects.
Then we arrive at a ‘‘meta confidence interval’’ for the employment effect of
[)36.5, )25.3] percentage points. This is fairly wide, but a hypothesis that all
AFP retirement is substitution from other retirement paths is clearly rejected.
On the other hand, a hypothesis that there is no substitution (50% points
reduction in employment) is also clearly rejected.
Table 4. Nonparametric estimates of the effect of AFP entitlement on the AFP)sample. Percentages

Observed Estimated effect

b c
AFP Non-AFP Unweighted Nearest neighbora Radiusa, Kernela, Intervala
Early retirement

Still in work 34.9 63.7 )28.8 [)31.9, )25.6] )32.2 [)36.5, )26.6] )30.1 [)33.4, )26.6] )30.0 [)34.1, )25.3] )30.4 [)34.7, )26.0]
Disabled 5.7 13.7 )8.0 [)10.0, )6.0] )6.9 [)10.7, )4.3] )7.5 [)9.9, )5.5] )8.5 [)11.2, )5.4] )8.3 [)11.2, )5.5]
Sick leave 5.3 8.4 )3.1 [)4.7, )1.4] )4.4 [)7.3, )0.5] )3.6 [)5.7, )2.0] )4.2 [)6.7, )1.5] )4.3 [)7.0, )1.7]
Other 4.1 14.2 )10.1 [)12.1, )8.2] )6.5 [)11.2, )4.6] )8.8 [)10.9, )7.3] )7.2 [)10.0, )5.2] )6.9 [)9.4, )4.5]
AFP 50.0 –
Observations 1939 1566 1939 1939 1939 1939 1939
Comparisons 1566 1237 1566 1566 1566

Confidence intervals in brackets.

a
Matching based on the propensity score. Bootstrapped confidence intervals in brackets (200 replications).
b
Radius: 0.1.
c
Gaussian kernel.
401
402 E. Bratberg et al.

5.3. Parametric estimates

Table 5 shows parametric estimates based on the multinomial logit model.

The regression results are reported in the appendix. Differently from the non-
parametric estimates, the control group here is used only for estimating the
model. The counterfactual is constructed by applying the estimated coeffi-
cients for the non-eligibles on the eligibles. Column 3 reports the estimated
effects. The estimates are quite similar to the unweighted differences in
Table 4. This reflects that the comparisons are made at the mean of the AFP-
group, and may indicate that differences in the distributions of the
conditioning variables are not properly accounted for. It is also possible to
use the logit estimates to compute the marginal effect of potential AFP
income on the outcome for an individual who is eligible. We find that the
effect of increasing AFP income by NOK 10 000 is –0.01 (evaluated at the
average probabilities, with Pwork = 0.35).14
To shed more light on the hypothesis of Section 2 that the incentive to
retire early is highest for low-income individuals, Table 6 splits the AFP-
sample by income level. The AFP-effect is then computed in the same way as
in Table 5. The observed fraction of AFP pensioners is about 9 percentage
points larger in the low-income group, in accordance with the prediction. The
counter-factual percentages (computed after re-estimating the model for each
sub-group) are somewhat less conclusive. On the one hand, substitution from
disability to AFP is larger for the low-income group; on the other hand, the
% substitution from work is larger for the high-income group (56.4 vs.
52.9%). However, these point estimates are uncertain.
Finally, we may combine the results from the parametric analysis with the
results using the propensity score to judge the AFP effects. The combined
confidence intervals (as it turns out, unaffected by the parametric estimates)
are Work: [-36.5, -25.3], Disability: [-11.2, -4.3], Sick leave: [-7.3, -0.5], Other:
[-11.2, -4.5]. To obtain ‘‘substitution rates’’, divide by 50% (the percentage
AFP-pensioners in the AFP sample). For instance, the substitution from
work is at least 50.6%, whereas the substitution from disability is at most
22.4% but at least 8.6%. As disability benefits are subject to medical
screening, the true number may be closer to the lower limit. Røed and
Haugen (2003) find less (virtually no) substitution between early retirement
and disability benefits. These results are not quite comparable due to different
samples, outcome definitions and methodology, however.15 We also find some

Table 5. Parametric estimates of the effect of AFP on the AFP sample. Percentages

Observed Without AFP Difference Difference as %

entitlementa of observed
AFP

Still in work 34.9 63.2 [60.9, 65.6] )28.3 [)31.2, )25.6] 56.6 [51.7, 62.0]
Disabled 5.7 14.4 [12.3, 16.3] ) 8.7 [)10.7, )6.4] 17.3 [12.9, 21.3]
Sick leave 5.3 9.2 [7.6, 10.8] )3.9 [)5.6, )2.3] 7.9 [4.6, 11.3]
Other 4.1 13.2 [11.6, 14.7] )9.1 [)10.7, )7.5] 18.3 [15.0, 21.5]
AFP 50.0 – – –

N=1939.
Bootstrapped confidence intervals in brackets (1000 replications).
a
Computed by using estimated logit coefficients from non-AFP sample on AFP sample.
Table 6. Transitions in the AFP sample by income group. Percentages

Part-time Low-incomea High-incomeb

OF WOT Diff. % of OF WOT Diff. % of OF WOT Diff. % of

AFP AFP AFP
Early retirement

Still in work 43.7 70.8 )27.1 63.8 28.0 57.7 )29.7 52.9 38.1 64.6 )26.5 56.4
Disabled 4.5 10.6 )6.0 14.2 7.3 19.9 )12.6 22.4 4.7 10.7 )6.1 12.9
Sick leave 5.1 6.5 )1.4 3.4 6.2 10.4 )4.3 7.6 4.6 8.7 )4.1 8.8
Other 4.2 12.1 )7.9 18.6 2.4 12.0 )9.6 17.1 5.7 16.0 )10.3 22.0
AFP 42.5 – – – 56.1 – – – 47.1 – – –
N 332 793 814

OF: Observed fraction; WOT: Without entitlement (computed by using estimated coefficients from non AFP sample, estimated by income group, on AFP sample).
a
Low-income: Working full time with income £ NOK 150000.
b
High-income: Working full time with income > NOK 150000.
403
404 E. Bratberg et al.

substitution from the long term sick (who also receive NIA benefits). Still, it
seems fair to conclude that by a conservative judgment, at least 50% of AFP
retirement is induced by the existence of the AFP scheme. As a final caveat,
we remind the reader that these results are based on private employees, so the
magnitudes of the effects are not necessarily directly extendable to public
servants.

6. Final remarks

The introduction of the AFP early retirement scheme for 66-year-olds in 1989
in many respects had the property of a natural experiment. Data from 1989
are not available, thus we are not able to make the most of this experiment.
We have argued, based on institutional circumstances as well as available
information in the MOTIPE data, that even for our time window this
property is retained. Therefore, we estimate the AFP effect, considering our
data as quasi-experimental, improving pure ‘‘experimental’’ comparisons by
non-parametric matching and a parametric estimator. Admittedly, this
conditioning might have been improved if more firm specific information
were available. Even so, we find the results quite convincing, indicating that
the economic incentives of the AFP early retirement scheme to a large extent
drive individuals out of the labor force – as predicted by theory.
Discussing our findings in light of the related literature reviewed above, we
first note that our analysis confirms that the economic incentives in the AFP
scheme induce early retirement. This corresponds to the results of Hernæs
et al. (2000) and Røed and Haugen (2003), and more generally to a bulk of
studies in various countries of the induced retirement effects of non-actuarial
early retirement programs.
Second, we find that the introduction of a voluntary early retirement
program to some extent may substitute for existing programs like disability
pensions and sick leave benefits. The ‘‘net’’ effect still is that at least 50% of
the AFP retirees (a conservative lower bound) would otherwise have been
employed. This quantitative assessment confirms and strengthens the finding
of Pedersen and Smith (1996) and Lilja (1996) that different sets of factors
explain exits to respectively voluntary programs and programs like disability
pensions. We note that the total cost of the AFP scheme for society is given
by the output forgone due to the net reduction in labor supply. This cost will
increase significantly over time as a consequence of larger cohorts reaching
the AFP eligibility age.16
Turning finally to policy implications, we know that a main policy
objective in almost all OECD economies including Norway, is to stimulate
labor supply in order to combat negative effects of ageing on growth and
social security financing. Clearly, this objective implies, among other things,
that economic policies should counteract the observed escalation of early
retirement. The main policy message of this paper is therefore to highlight the
need for improvements of the economic incentives to stay longer in the
workforce. This calls for social security reforms that increase the early
retirement schemes’ marginal degree of actuarial fairness.
Early retirement 405

Appendix

Table A1. Logit estimates for the propensity score

Coef. Std. Err.

Male 0.107 0.109

Experience 0.017 0.014
Sickdays 0.001 0.001
Unempdays 0.001 0.001
(Unempdays)2 )2.660E)06 2.680E)06
Unempdays*Sickd. 1.140E)05 1.170E)05
Petrol 1.659 0.390
Manufacturing 2.033 0.221
Construction 1.023 0.255
Retail 0.292 0.216
Transport 0.158 0.235
Finance )0.360 0.246
Service )0.592 0.233
Region1 )0.857 0.155
Region2 )0.110 0.163
Region3 )0.566 0.157
Constant )0.409 0.387
N 3505
Loglikelihood )2013.1
Pseudo R-squared 0.165

Dependent variable is T = 1 if entitled to AFP, 0 otherwise.

Table A2. Distribution of the propensity score (p)

T=0 T=1 Total

Inferior bound of p
0.11 90 19 109
0.20 774 335 1109
0.40 246 208 454
0.50 185 288 473
0.60 52 74 126
0.70 24 37 61
0.75 88 313 401
0.80 107 665 772
Mean (S.D.) of p 0.44 (0.19) 0.65 (0.21) 0.55 (0.23)
Total 1566 1939 3505

Number of blocks determined such that mean propensity score is not different for treated and
controls in each block.
Within each block means of the conditioning variables are tested for the balancing property.
406 E. Bratberg et al.

Table A3. Multinomial logit results

AFP Non-AFP

Coef SE z Coef SE z

Income (state specific) 0.039 0.011 3.500 0.028 0.010 2.740

Disability
Male 0.442 0.350 1.260 )0.436 0.216 )2.020
Experience )0.001 0.040 )0.030 0.019 0.027 0.700
Unempdays 0.000 0.001 0.160 )0.003 0.002 )1.740
Sickdays 0.007 0.002 4.760 0.005 0.001 3.530
Married 0.152 0.279 0.540 0.381 0.205 1.860
Fulltime 0.528 0.335 1.580 0.752 0.220 3.410
Transp. 0.228 0.235 0.970 )0.683 0.207 )3.290
Service 0.078 0.347 0.230 0.199 0.199 1.000
Constant )2.724 0.937 )2.910 )2.365 0.664 )3.560
Long term sick leave
Male )0.338 0.316 )1.070 )0.083 0.276 )0.300
Experience 0.022 0.038 0.570 )0.029 0.032 )0.910
Unempdays )0.003 0.002 )1.360 0.000 0.001 )0.170
Sickdays 0.006 0.002 3.320 0.009 0.001 6.010
Married 0.372 0.293 1.270 0.286 0.251 1.140
Fulltime 0.440 0.325 1.350 0.706 0.270 2.610
Transp. )0.013 0.244 )0.050 0.009 0.244 0.040
Service )0.109 0.359 )0.300 )0.170 0.273 )0.620
Constant )2.880 0.902 )3.190 )2.159 0.753 )2.870
Other
Male 0.895 0.414 2.160 )0.242 0.207 )1.170
Experience )0.026 0.044 )0.600 0.026 0.026 1.000
Unempdays )0.001 0.002 )0.630 )0.002 0.001 )1.570
Sickdays 0.001 0.003 0.350 0.002 0.002 1.310
Married )0.290 0.288 )1.010 0.102 0.188 0.540
Fulltime 0.277 0.365 0.760 0.387 0.201 1.930
Transp. 0.647 0.271 2.390 0.382 0.156 2.450
Service 0.654 0.357 1.830 )2.278 0.620 )3.670
Constant )2.143 0.970 )2.210 )0.242 0.207 )1.170
AFP
Male )0.027 0.159 )0.170
Experience 0.098 0.021 4.660
Unempdays 0.001 0.001 1.220
Sickdays 0.004 0.001 3.300
Married 0.064 0.132 0.490
Fulltime 0.226 0.156 1.450
Transp. )0.120 0.119 )1.010
Service )0.428 0.177 )2.420
Constant )2.211 0.504 )4.390
N 1939 1566
Loglikelihood )2193.5 )1580.9

Reference state: Still in work.

Potential income if working assumed to be same as observed 1993 income. Potential income in the
‘‘other’’ state approximated by using average income of those observed in that state, by gender.
Early retirement 407

Endnotes
1
See for example the surveys of OECD (1998a, 1998b) and Gruber and Wise (1997). Börsch-
Supan (2000) considers the incentives in European pension systems to retire early.
2
As highlighted by OECD (1998a) the tendency to liberalize the entitlement conditions to
disability and unemployment programs during recessions has been widespread all over
Europe.
3
For more details about the AFP program see the Norwegian government white paper NOU
1998:19. A useful and compact survey of the institutional settings in Norway is offered by
Hernæs et al. (2000).
4
Through various regulations in the Norwegian tax-system, the taxation of retirement income
is somewhat milder than for regular labor income.
5
In order to understand the implication of this ‘‘free earnings point’’ property of the AFP
scheme, we may simply imagine an individual aged 64, who considers early retirement and has
37 years of labor income above the minimum level. At the outset the missing 3 years (in
comparison to the 40 years requirement for a full supplementary pension) implies a 3/40
reduction in the supplementary pension level. However, the AFP scheme takes a projection of
the labor income in the three early retirement years into account. This implies a full
supplementary pension during both the early retirement years as well as beyond 67. As noted
by Hernæs et al. (2000), AFP does not imply any loss in the pension level after age 67 provided
that labor earnings would not have increased from the level immediately prior to early
retirement.
6
Detailed calculations of net and gross replacement rates for the AFP scheme are presented in
the government white paper NOU 1994:2.
7
Salvanes and Førre (2003) consider job creation and destruction in different age cohorts. It
turns out that net job creation in the old cohorts is negative because job creation is negligible
and more than offset by the tendency to withdraw from the labor force.
8
It is well known that the MNL suffers from the independence of irrelevant alternatives (IIA)
property which may bias the results. The emphasis here is not on interpretation of the
coefficients, but on extrapolation to different background characteristics as indicated by Eq.
(9). Even so, we cannot claim that the IIA assumption is unproblematic.
9
Firms that are recorded as not offering the AFP option in this file could have been doing so
before 31.12.95. Therefore the AFP group was supplemented by identifying firms where
employees retired on the AFP scheme during 1994 or 1995, and re-classifying them as AFP
firms if not already belonging to that group. 109 of a total 1472 AFP firms were identified in
this fashion. There were 1360 non-AFP firms.
10
The employee register gives start and stop dates and a coarse indicator of employment (4–19,
20–29, 30+ hours week), but not exact information on working hours. The variable is missing
for 1992 in our data.
11
Only 17 individuals were in the unemployment register at 66½ , however some individuals in
the disability and sick leave categories had previously been unemployed.
12
On average, AFP and disability pensioners retired 248 and 286 days after their 64th birthday,
respectively. 51% (37%) of the AFP (disability) retirees retired within 6 months, and 70%
(65%) within a year.
13
The matching estimators were computed using code written by Becker and Ichino (2002).
14
The marginal effect of an increase in the income in state k on state j is (1(j = k)Pj – PjPk)c,
where c is the coefficient on income, see e.g., Greene (2003).
15
Røed and Haugen consider public servants as well as private employees. Their early retirement
category includes AFP, as well as withdrawal from the labor force for other reasons than
disability, sickness or unemployment.
16
As explained in Sect. 2.1 the government’s direct financial outlays to the AFP scheme is limited
to 40% of the benefits to AFP pensioneers aged 64, 65 and 66. ‘‘Back of the envelope’’
calculations based on approximately 10500 AFP pensioneers during the mid-1990s and an
average AFP benefit equal to 113000 NOK suggest a gross outlay equal to 475 millions NOK
and a net outlay (adjusting for substitution between transfer schemes) roughly equal to the
half. These figures do not include the financial costs of ‘‘free’’ earnings points to regular
supplementary pension and the implied loss of tax revenues caused by individuals’ exit from
the labor force. Attempts to estimate the future total cost for the society caused by the net
408 E. Bratberg et al.

reduction in labor supply in all AFP cohorts call for numerical simulations that capture the
expected ageing of the population. We note that simulations in Thøgersen et al. (1997) suggest
that this cost will amount to a 15400 NOK reduction in the net lifetime income of a
representative individual.

References
Becker SO, Ichino A (2002) Estimation of Average Treatment Effects Based on Propensity
Scores. Stata Journal 2: 358–377
Börsch-Supan A (2000) Incentive Effects of Social Security on Labor Force Participation:
Evidence in Germany and across Europe. Journal of Public Economics 78: 25–49
Dahl SÅ, Nilsen ØA, Vaage K (2000) Work or Retirement? Exit Routes for Norwegian Elderly.
Applied Economics 32: 1865–1876
Dehejia RH, Wahba S (2002) Propensity Score-Matching Methods for Nonexperimental Causal
Studies. Review of Economics and Statistics 84: 151–161
Fehr H, Sterkeby WI, Thøgersen Ø (2003) Social Security Reforms and Early Retirement. Journal
of Population Economics 16: 345–361
Greene WH (2003) Econometric Analysis, 5th ed. Prentice Hall, New Jersey
Gruber J, Wise D (1997): Social Security Programs and Retirement around the World. NBER
Working paper no. 6134, National Bureau of Economic Research, Cambridge, MA
Heckman JJ, Ichimura H, Todd P (1998) Matching as an Econometric Evaluation Estimator,
Review of Economic Studies 65: 261–294
Heckman JJ, Smith J (1995) Assessing the Case for Social Experiments. Journal of Economic
Perspectives 9(2): 85–110
Hernæs E, Sollie M, Strøm S (2000) Early Retirement and Economic Incentives. Scandinavian
Journal of Economics 102: 481–502
Krueger A, Pische JS (1992) The Effects of Social Security on Labor Supply: A Cohort Analysis
of the Notch Generation. Journal of Labor Economics 10: 412–437
Lilja R (1996) Microeconomic Analysis of Early Retirement in Finland. In: Wadensjö E (ed) The
Nordic Labour Markets in the 1990s. Elsevier Science Publishers, Amsterdam
Lumsdaine RL, Mitchell OS (1999) New Developments in the Economic Analysis of Retirement.
In: Ashenfelter O, Card D (eds) Handbook of Labor Economics, Vol. 3C. Elsevier Science
Publishers, Amsterdam
NOU 1994:2 Fra arbeid til pensjon (‘‘From work to retirement’’). Norwegian Government White
Paper, Oslo
NOU 1998:19 Fleksibel pensjonering (‘‘Flexible retirement’’). Norwegian Government White
Paper, Oslo
OECD (1998a) The Retirement Decision, Chap. VI. In: OECD Economic Outlook 1998. OECD,
Paris
OECD (1998b) The Retirement Decision in OECD Countries. Working Paper No. 1 on
Macroeconomic and Structural Policy Analysis (ECO/CPE/WP1(98)2), OECD. Paris
Pedersen PJ, Smith N (1996) Duration Analysis of the Decision to Retire Early. In: Wadensjö E
(ed) The Nordic Labour Markets in the 1990s. Elsevier Science Publishers, Amsterdam
Rosenbaum P, Rubin D (1983) The Central Role of the Propensity Score in Observational
Studies for Causal Effects. Biometrika 70: 41–55
Røed K, Haugen F (2003) Early Retirement and Economic Incentives – Evidence from a Quasi-
Natural Experiment. Labour (forthcoming)
Salvanes KG, Førre SE (2003) Employment Effects of Trade and Technical Change: Evidence
from Norway. Economica 70: 293–330
Thøgersen Ø, Andersen C, Rogdaberg G, Sannarnes JG (1997) Fleksibel pensjonsalder,
folketrygdens finansiering og fordeling av velferd mellom generasjoner (‘‘Early retirement,
social security and intergenerational distribution’’). Report 48/97, SNF - Foundation for
research in economics and business administration, Bergen