PENMICRO

Monitoring pension developments through micro socio-

economic instruments based on individual data sources:
feasibility study

Final Report
for
The European Commission
Employment, Social Affairs and Equal Opportunities
DG EMPL E4 Unit

Submitted by:

TARKI Social Research Institute (Hungary)

Prepared by:
Róbert I. Gál
András Horváth
Gábor Orbán
in collaboration with
Gijs Dekkers (FPB, Belgium)
Table of contents

Executive summary .................................................................................................................... 3

1. Introduction ............................................................................................................................ 5
2. Types of pension models........................................................................................................ 7
2.1 Standard models ............................................................................................................... 8
2.1.1 Cohort models ........................................................................................................... 8
2.1.2 Typical agent models .............................................................................................. 11
2.2 Microsimulation models................................................................................................. 12
3. Approaches to pension modelling ........................................................................................ 18
3.1 Exclusive use of standard models .................................................................................. 18
3.1.1 Imported models customised to special needs ........................................................ 19
3.1.2 Own developments.................................................................................................. 21
3.1.3 Experiences of introducing microsimulation to the modelling toolkit.................... 24
3.2 Micro datasets ................................................................................................................ 25
3.3 Utilisation of microsimulation models ........................................................................... 29
4. An inventory of pension projection models ......................................................................... 33
References ................................................................................................................................ 55
Annex: The PENMICRO questionnaire................................................................................... 57

2
Executive summary

1. Throughout this report we differentiate between standard (cohort or typical agent) models
and microsimulation models.
Cohort models start from current cross-sectional information on labour market position and
social security contributions of various social groups, most frequently cohorts, which are
sometimes further disaggregated by gender and some labour-market attributes, including
demographic features such as marital status, and level of education. Most cohort models are
partial equilibrium deterministic models but there are also special overlapping generation
general equilibrium cohort models in use incorporating the economic behaviour of house-
holds, firms and pension funds. Typical agent models draw typical histories of fictitious indi-
viduals (i.e. the unemployed or minimum wage earners as such, etc.), based on which the
level of benefits is calculated. This method has a strong longitudinal element, but the cross-
sectional view is rather ad hoc, especially as transitions across the fictitious sub-groups that
form the basis of this model are usually not accounted for.
Microsimulation models simulate changes in a sample of a large number: thousands and
sometimes hundreds of thousands of individuals. In their simplest form, microsimulation
models are static, i.e. they compare two states of the world, or two different institutional set-
tings and simulate “overnight effects”. In contrast, dynamic microsimulation models include
time, and hence introduce demographic ageing. There are two approaches to introducing age-
ing in a dynamic microsimulation model. The first is dynamic microsimulation with static
ageing, in which cross-section characteristics are updated by exogenous future aggregate data,
so that time is traced as a series of world states. In practice, dynamic microsimulation models
with static ageing first reweigh the individual cases in order to adjust the sample to exogenous
projected demographic and labour market developments. Then, as a second step the resulting
aggregate results are adjusted by some exogenous developments, such as economic growth.
This reweighing-updating process resembles what standard cohort models achieve, except that
for microsimulation the number of cases reweighed and updated is much larger. Dynamic
microsimulation models with dynamic ageing go further, by building up complete life histo-
ries of each individual in the dataset.
2. Approaches of Member States to pension modelling can be best classified by their use of
microsimulation instruments and their modelling capacity. This implies two distinct fields of
capacity building:
- establishing or improving in-house modelling activity in the administration, and
- creating or extending micro datasets that can be used in microsimulation.
The latter appears to be more urgent but also more demanding of the administration.
3. We used the following taxonomy for categorising national approaches (countries falling in
the respective classes are in parentheses):
ƒ Only standard (cohort, typical agent) model but no micro-simulation model
ƒ Import model from international or foreign agency customised to special needs
(Bulgaria, Cyprus, Latvia, Lithuania, Malta, Portugal)
ƒ Model developed in-house
ƒ No steps towards microsimulation (Estonia, Greece, Ireland, Poland, Slovakia,
Slovenia)
ƒ Steps towards microsimulation (Czech Republic, Spain, Hungary, Austria)
ƒ Utilisation of microsimulation model
ƒ Microsimulation model as complement
ƒ Modelling work contracted out (Germany, Italy)

3
 ƒ Modelling work in-house (Belgium, Denmark, Luxemburg, Netherlands, Fin-
land)
ƒ Microsimulation as primary modelling instrument (France, Sweden, United King-
dom)
4. Several remarks need to be added to the above typology.
First, standard models can differ significantly from one another in their level of so-
phistication so less developed models, and imported models in particular, can be worth further
improving.
Second, among microsimulation models the major step is introducing time into the
model that is to move from static to dynamic.
Third, applying microsimulation models as complements is not necessarily a sign of
the lack of administrative capacity but it could be an optimal use of resources available for the
administration.
5. We found five successful ways of introducing micro datasets to microsimulation models
starting from the minimal requirements to the ideal cases. These are the following (countries
falling in the respective classes are in parentheses; in case of more models by country, the
name of the model is also added):
- Using social insurance administrative data only (Hungary, Sweden-MiMESIS)
- Matching social insurance administrative data with tax records and/or census data
(Spain, Netherlands-SADNAP, Austria, Finland)
- Using large household surveys (France-DESTINIE, Netherlands-MICROS, MI-
DAS (various countries))
- Matching social insurance administrative data with household surveys (France-
PRISME, Luxemburg, United Kingdom)
- Matching numerous administrative datasets (Belgium, Denmark, Sweden-SESIM)

4
1. Introduction

TARKI Social Research Institute was awarded a contract by the European Commission to
prepare a feasibility study on “Monitoring pension developments through micro socio-
economic instruments based on individual data sources”. From the onset the project was de-
signed so as to
- Inform on the state of development of instruments used by Member States to moni-
tor life time earnings and related acquisition of pension rights
- Provide descriptions of available tools, such as data sets and models of prospective
pension outcomes based on individual data sources
- Identify and characterise tools, in particular where they already exist and where
they could be further developed (notably needs for adequate data sources, need for
building models).
- Analyse and describe both the administrative and survey data used for the prospec-
tive models of pension outcomes across all Member States (where such instruments
exist) and elaborate a classification of instruments built to assess future develop-
ments of pension benefits, based on individual information
- Characterise the various results available and review them as well as those likely to
be derived through such instruments
This design allows the research team to provide a description of the available pension model-
ling instruments and the datasets behind them for each Member State; to describe the out-
comes of these instruments (focusing mostly on Member States applying the most developed
tools); and to present an evaluation of similarities and differences between the national mod-
els and approaches, all along the lines of the Invitation to tender No VT/2007/115 and the
resulting Study Service Contract VC/2007/0443.
The research team used the following sources of information. In the exploratory phase we
reviewed and analysed the report on Pensions Schemes and Projection Models in EU-25
Member States by The Economic Policy Committee and Directorate-General for Economic
and Financial affairs; responses to the first questionnaire of the Indicator Sub-Groups of the
Social Protection Committee (ISG); presentations of national delegates at the ISG workshop
in Madrid in January 2008; as well as corresponded with the national delegates to the ISG and
searched the internet. This phase resulted in an interim report, which contained a detailed
questionnaire and a first inventory of instruments. In the second phase we turned again to the
national delegates requesting them to complete the questionnaires and the relevant parts of the
inventory. Finally, in the third phase, we used the collected material to write the comparative
sections.
The research team is indebted to the national delegates of the ISG for their contribution.
There are also some caveats to be made in this introduction. The scope of the study is instru-
ments in use by administrations. Academic or business models are left aside as long as they
are not part of the administrative work. We covered only those models that were named by the
national delegates or were clearly identifiable as administrative. Also, the information col-
lected had its limits due to the number of no-responses.

The final report is organised into three sections. Section 2 offers a typology of pension model-
ling instruments and uses this structure to give a comparative view of models applied by
Member State administrations. In Section 3, it is not models but administrations that are com-
pared as to their approach to the use of micro data and microsimulation models in pension

5
projections. The section also contains two subsections on the experiences of introducing mi-
crosimulation to the modelling toolkit and on the successful ways of putting micro datasets to
work in pension modelling. Finally, Section 4 gives an all-European inventory of administra-
tive pension models.

6
2. Types of pension models

In this section we describe, classify and characterise modelling instruments applied by ad-
ministrations of Member States. The unit of analysis is the model. This is in contrast with
Section 3, where the unit of analysis will be the administration, which may use more than one
model for pension projections. In cooperation with the national delegates to the Indicator Sub-
Groups of the Social Protection Committee we identified 44 such tools in 26 countries (see
Table 1). In the table we present a list of the models by country. In some cases, in the absence
of any official version we ascribed our own names to the models (such as German Pension
Model). The emphasis is put on comparison of model properties. Detailed descriptions of the
individual models can be found in Section 4, the inventory of administrative modelling in-
struments.

Table 1: Pension modelling instruments applied by governments of Member States

Belgium MALTESE; MEP; MIDAS_BE; MIMOSIS
Bulgaria ILO PENS (BG)
Czech Republic Czech Pension Model
Denmark Danish Pension Model; LAW
Germany AVID; German Pension Model
Estonia Estonian Long Term Pension Budget Model
Ireland Irish Pension Model
Greece Greek Pension Model
Spain Spanish Pension Model
France DESTINIE; PRISME
Italy CAPP_DYN; CeRP models; RGS
Cyprus ILO PENS (CY)
Latvia Latvian Pension Model
Lithuania PRISM
Luxemburg LuxMod; REDIS; SOBOLUX
Hungary Hungarian Pension Models; NYIKA
Malta PROST (MT)
Netherlands GAMMA; MICROS; SADNAP
Austria Austrian Applied Projection Models; Austrian Microsimulation Model
Poland FUS07
Portugal ModPensPor
Romania na
Slovenia SIOLG 1.0
Slovakia PROST (SK), MAJA
Finland Finnish Centre for Pensions models
Sweden MiMESIS, SESIM
United Kingdom PENSIM2

We apply taxonomies by Zaidi and Rake (2001), Dupont, Hagneré and Touzé (2003), Dek-
kers (2007), Harding (2007) and European Commission (2007) in order to classify the instru-
ments in use by European governments.

7
2.1 Standard models
2.1.1 Cohort models
The majority of models under study in this report can be best classified as standard cohort
models. Instruments of this type start from current cross-sectional information on labour mar-
ket position and social security contributions of various social groups, most frequently co-
horts, which are sometimes further disaggregated by gender and some labour-market attrib-
utes, including demographic features such as marital status, and level of education. The in-
formation set, which consists of group-averages, is sometimes extended to include retrospec-
tive data on contributory or wage history. Current information is projected then into the future
subject to assumptions on demographic and labour market developments. These parameters in
effect re-weigh the group-averages. A partial equilibrium concept is usually applied in order
to assure consistency among the aggregates resulting from the re-weighing process and some
exogenous macroeconomic variables. In addition, often some mechanisms are applied in order
to ensure internal consistency among the parameters of the projection. Standard cohort models
have the advantage, as compared with the typical agent models (see below), that it is easier to
obtain aggregate conclusions, such as future revenues or expenditures of the pension system.
It requires special care from the modeller to secure consistency between projections on wages
and those of employment levels and wages. Translating exogenous cross-section labour force
assumptions into cohort average (longitudinal) career characteristics (length of service, bene-
fit base) is often a critical point in these cohort models.
The main issues related to cohort models are the creation of sub-groups (usually cohorts and
gender groups and sometimes even further disaggregated sub-groups), the assumptions re-
garding their future prospects, and the mechanism applied to reweigh these sub-groups in
light of expected future developments and the updating process which warrants for the consis-
tency among parameter settings, and that of aggregate results with exogenous macroeconomic
and financial variables. Some other questions are also introduced to the form in order to cap-
ture potential extensions, such as the option of parallel work and retirement (more specifi-
cally, gradual retirement), the availability of retrospective data on the sub-groups and applica-
tion of various empirical models to derive behavioural responses in savings, fertility and la-
bour supply.

Input
Standard cohort models covered in this report start out from cohort aggregate figures of ad-
ministrative datasets, broken down by benefit type and pension scheme. The Italian CeRP-
SAM takes into account also geographical differences (North vs. South), while the Danish
model provides a breakdown by ethnic origin. The Irish model treats the self-employed sepa-
rately.
The macro environment is exogenous in the standard models applied by EU administrations.
Given that no feedback from the pension system parameters is captured, behavioural re-
sponses are absent altogether from the analysis. The internal consistency of macro variables
such as wage growth, employment and inflation is usually ensured by the general principle
that the labour share of GDP is held constant. This common assumption originates from the
AWG comparative exercise on the economic consequences of ageing. In general, most stan-
dard models, as well as most micro models reviewed in this report, assure consistency of
macro variables by an external mechanism, not endogenously. Exceptions are special standard
models with more sophisticated equilibrium mechanisms (see the subsection on special cohort
models below).

8
Demography is exogenous in the standard models covered here, with the exception of the
Lithuanian version of PRISM, which is able to generate a demographic projection of its own.
Generally, the same mortality rates are applied across all social groups. Exceptions are the
Hungarian standard model, Poland's FUS07 and the standard model of the Finnish Centre for
Pensions, which apply different mortality rates of certain types of benefit recipients (the dis-
abled, in particular). Moreover, in the latter instrument, for persons drawing an old-age pen-
sion a high pension is connected with a low mortality risk when age and gender are standard-
ised.
In all cases, aggregate labour market developments are exogenous as well. This does not ex-
clude the possibility of taking into account other empirical phenomena impacting the labour
market. In the Italian RGS, participation rates are affected by the following three factors: i)
the direct and indirect effects brought about by the evolution of enrolment rates, the latter
through changes in educational achievements, ii) the fulfilment of the eligibility requirements
to be entitled to a pension, taking also into account the evolution of workers distributed by age
and contribution years and iii) changes in the labour market equilibrium caused by the de-
crease in the working age population. In the Hungarian standard models, the participation
rates are adjusted outside of the model to incorporate views on school enrolment (lower par-
ticipation in younger cohorts today vs. higher participation of older cohorts in the decades to
come).

Modelling
All standard models are deterministic by definition. They address the issue of fundamental
uncertainty regarding the underlying assumptions and other types of model uncertainty by
relying on sensitivity analyses and robustness checks, in order to test how unstable the results
are when macro assumptions, demographic scenarios, age-profiles or future labour market
scenario changes.
Most standard models under review here differentiate with respect to sex, age and benefit
type, but some go further. For instance, to address the issue of immigration, the Danish in-
strument, as mentioned above, was designed to include dimensions such as ethnic origin. In
order to take account of peculiarities of the Luxembourg labour market (high proportion of
migrant workers), LuxMOD was designed to include dimensions such as country of origin or
employment status (beyond the general breakdown by age, sex and benefit type). The Belgian
MALTESE and the Czech model distinguish between employed and self-employed and the
latter also on the basis of income ranges (10 per gender). SOBULUX, another instrument ap-
plied by the administration in Luxemburg, differentiates blue-collar workers, white-collar
workers and civil servants. Households are not a unit of analysis in any of the models under
study in this section.
Another important issue is whether the entry of new pensioners into the existing pool is ex-
plicitly modelled and how. This is highly pivotal given that the simulation of different policy
scenarios is only possible if rules applying to new pensioners (i.e. the raising of the retirement
age or a change in the pension formula) can be introduced to the model in a transparent way.
Even more importantly, the effects of past measures, which take time to feed through the sys-
tem, can only be captured explicitly in this way. Naturally, this issue is tackled by instru-
ments, which track individuals through their life path, but this is not the case for most stan-
dard models. Most models, however, tend to have manually driven retirement. This means
that retirement is modelled by the adjustment of retirement age profile based on changes in
participation rates of older cohorts (as in the Estonian instrument’s case), changes in retire-
ment age legislation (in Hungary), etc. An exception is the Portuguese model, in which de-

9
termining the flow of new retirees is performed in a simplified way: the proportion of a cohort
becoming eligible for pensions is assumed to stay constant from the base year onwards. The
average length of contributions is also held constant, without incorporating projected changes
in the labour market. New disability pensioners are projected either assuming an evolution of
the stock (as in the Irish model), or explicitly. In the latter case, most models covered in this
study build upon past experience, holding the rate of new disability pensioners constant.
MALTESE takes a more sophisticated approach by calculating entry probabilities into the
disability scheme from the exogenous labour force projection (number of potentially active
people).
A disadvantage of the cohort-based approach is that in cases where there are non-linear rela-
tionships in the pension systems (i.e. a minimum number of service years required for pension
eligibility, minimum pension or different rates of benefit indexation based on benefit amount),
the cohort averages may be distorted. In these cases the information on the distribution influ-
ences the updating of the average value. Most models do not address this issue or do so only
partially. The Italian RGS model is cohort-based and follows an indirect method and supple-
ments the mean value with an index of variability (the variation coefficient) and a distribution
function, whenever the influence of the distribution is significant. Whilst the aforementioned
example shows that some effort is made to overcome the distortion problem, the limits of the
standard framework are clear.

Output
The most important output of the standard models considered here are aggregate revenues and
expenditure and the number of contributors/beneficiaries to the schemes. Sustainability indi-
cators produced in most cases are the deficit of the pension system (financing requirement
beyond that provided by contributions) and cumulative indicators, such as the implicit debt of
the system.
Modelling instruments are important to compare through the concept of replacement rate they
apply. Theoretically, benefit levels are usually compared to wages (replacement rates) or to
contributions (rate of return on contributions). Standard models under study in this report pro-
duce aggregate empirical replacement rates; households and individuals are not modelled. In
these models, heterogeneous outcomes across the population, particularly poverty and ine-
quality are not reflected in the projection outputs. Specifically, in most standard models cov-
ered here, the distribution of pension benefits cannot be obtained, therefore conclusions on
adequacy and/or inequality cannot be drawn. An exception may be the Lithuanian PRISM,
which can operate as an “average person parameter” model or as a distributional model. In
this case new retirees are assigned a bivariate normal distribution of lifetime wages and of
pension service. Czech old-age benefit projections are disaggregated on the basis of income
ranges; this feature allows it to project the share of new retirees under the poverty line. Rate
of return measures are provided by SOBULUX and CeRPSAM, but stock measures such as
pension wealth are not computed in the standard models under study in this report.

Special cohort models

Most cohort models are partial equilibrium deterministic models, though the meaning of ‘par-
tial’ may differ from one model to the other. For instance, as explained above, consistency
between the growth rate of the workforce, productivity and GDP growth is almost always
ensured through the assumption of a constant share of labour in total national income. How-
ever, other types of ‘endogeneities’ are more difficult to come by and are mostly incorporated

10
indirectly. The Dutch GAMMA and the Slovenian SIOLG 1.0 are special overlapping genera-
tion general equilibrium cohort models incorporating the economic behaviour of households,
firms and pension funds. LuxMOD is a computable general equilibrium model based on ag-
gregate administrative data, which covers all social security schemes but does not simulate the
acquisition of pension rights.
Whilst they incorporate less detail on the pension system itself (e.g. GAMMA relates the de-
velopment of first pillar pension expenditure to two factors: productivity in the economy and
the number of people over the age of 65), these models are better equipped to carry out wel-
fare analysis of policy reforms. Also, given that agents are rational, forward looking and op-
timise in a consistent microeconomic framework, internal consistency is better ensured.

2.1.2 Typical agent models

Another branch of models is the typical agent model. Instruments of this type draw typical
histories of fictitious individuals (i.e. the unemployed or minimum wage earners as such,
etc.), based on which the level of benefits is calculated. This method has a strong longitudinal
element, but the cross-sectional view is rather ad hoc, especially as transitions across the ficti-
tious sub-groups that form the basis of this model are usually not accounted for. This ap-
proach yields a sophisticated estimate of replacement rates, given that this type of model usu-
ally contains the country-specific institutional setting with all its non-linearities, ceilings etc.
In addition, the acquisition of pension rights is properly tracked, as full life histories are avail-
able. Typical agent models are well suited to assess incentives in the pension system to work
longer as well as to compare the relative generosity of schemes across various life paths
and/or countries. The main challenge of this approach is to come up with a fiscal projection,
given that it is incapable of weighing the fictitious groups of individuals in a way that would
reflect the composition of society. For the same reason, it is not possible to analyse distribu-
tional effects of policy alternatives. Also, this “typical agent” method does not incorporate a
model for individual behaviour, e.g. the choice of the year of retirement, or macroeconomic
repercussions (e.g. higher contribution rates reduce labour force participation).

Typical agent models may differ in the main features and life-path characteristics of typical
agents, the motivations of choosing these particular fictitious actors, the way their life-paths
are set, and the mechanism that ensures consistency between core scenarios/assumptions. In
addition, there are different approaches to aggregating the results derived from typical life-
paths as non-exhaustive components and the empirical models used to derive behavioural
responses of typical agents to exogenous effects.

Among the models reviewed MEP (Microeconomic Pension Model) of Belgium is the only
genuine typical agent standard simulation model. It is based on the option value approach.
Four hypothetical private sector employees (male and female blue and white collar workers)
are modelled, who are deemed to be representative for the purpose of examining actuarial
neutrality of Belgian first pillar pensions. Individuals are assumed to be either single persons
or the sole income earners in the household. It is assumed that the four typical employees
were born in 1940 and entered the labour market at the age of 20. MEP calculates the flow of
expected future benefits until death for every year in which the individual is eligible for the
pension benefit. As outputs, it produces individual replacement rates, changes in pension
wealth, implicit taxes and option values.

11
The Greek Pension Model also performs typical agent simulations, which produce individual
theoretical replacement rates as their output. Simulations incorporate cases where people have
just entered employment or social insurance or will join the labour force in the near future or
became insured for the first time only a few years earlier. The fund modelled covers about 55
percent of all private sector employees.
The Czech and the Lithuanian model, as well as the versions of PROST operated by Slovakia
and Malta have a special feature in that they are able to run both as typical agent models and
as standard cohort models. The CeRP family of pension models also includes a typical agent
model, which is capable of computing pension benefits and money’s worth measures for a set
of representative workers according to the rules of the legislation in effect. By aggregating
individual results, CeRP evaluates the impact of alternative pension rules on pension expendi-
ture.

2.2 Microsimulation models

Microsimulation models simulate changes in a sample of a large number: thousands and
sometimes hundreds of thousands of individuals. Information on the sample can be collected
in surveys or by the data processing of various government agencies such as the tax authority
or the social security administration. They can be cross-sectional, if individuals appear in the
sample only once, or panel, if individuals can be followed over a period of time. Some sam-
ples include information only on the individual; some also contain data on other household
members. Administrative data have some obvious limitations compared to survey data, be-
cause administrative data are not collected for research purposes and thus usually lack some
information that would be required. Questionnaires for survey data may be designed specifi-
cally for defined research purposes. On the other hand, administrative data is in general more
detailed, complete, timely, regular and accurate than survey data.
In their simplest form, microsimulation models are static, i.e. they compare two states of the
world, or two different institutional settings and simulate “overnight effects” (Dekkers 2007).
In contrast, dynamic microsimulation models include time, and explicitly model life paths of
individual units as they age. Zaidi and Rake (2001) point out that while static models do not
attempt to incorporate behavioural changes, dynamic models are concerned to take on board
behavioural responses as well as simulating the policy environment.
Zaidi and Rake (2001) distinguish two approaches to introducing ageing in a dynamic mi-
crosimulation model. The first is dynamic microsimulation with static ageing, in which cross-
section characteristics are updated by exogenous future aggregate data, so that time is traced
as a series of world states. In practice, dynamic microsimulation models with static ageing
first reweigh the individual cases in order to adjust the sample to projected demographic and
labour market developments. Then, as a second step the resulting aggregate results are ad-
justed by some exogenous developments, such as economic growth (Harding 1996, Dekkers
2007). This reweighing-updating process resembles what standard cohort models achieve,
except that for microsimulation the number of cases reweighed and updated is much larger.

12
Table 2: Simulation properties of pension microsimulation models used by European administrations
Country Type of mi- Cross-sectional vs. Ageing process Initial database Database match- Information on the
crosimulation cohort / longitudi- ing household
nal
AMM1 Austria ? ? ? administrative yes no
AVID Germany dynamic longitudinal dynamic sample of adminis- yes y
trative
CAPP_DYN Italy dynamic cross-sectional dynamic survey no y
DESTINIE France dynamic cross-sectional dynamic survey no y
FCP2 models Finland ? ? ? administrative yes no
LAW Denmark static - - administrative yes y
MICROS Netherlands dynamic cross-sectional static and dynamic survey ? ?
MIDAS Belgium dynamic cross-sectional dynamic survey yes y
MiMESIS Sweden dynamic cross-sectional dynamic sample of adminis- yes ?
trative
MIMOSIS Belgium static - - sample of adminis- no no
trative
NYIKA Hungary dynamic cross-sectional dynamic administrative yes no
REDIS Luxemburg static - - administrative yes y
PENSIM2 United Kingdom dynamic cross-sectional dynamic sample of adminis- yes y
trative
PRISME France dynamic cross-sectional dynamic administrative no no
SADNAP Netherlands dynamic cross-sectional dynamic administrative yes y
SESIM Sweden dynamic cross-sectional dynamic sample of adminis- yes y
trative
1
Austrian Microsimulation Model; under construction.
2
Finnish Centre for Pensions models.

13
Dynamic microsimulation models with dynamic ageing go further, by building up complete
life histories of each individual in the dataset. Within this latter group, Dekkers (2007) differ-
entiates between cross-sectional models and cohort/longitudinal models as being two major
directions. In the former, individuals, who are the basic unit of modelling, are moved ahead in
time one-by-one a period, while each of their attributes is updated. In the latter an entire indi-
vidual life cycle is followed through from birth to death before the model takes in the next
individual. The advantage of cross-sectional models is that they allow for interactions be-
tween individuals, such as a marriage or death of a partner, in a straightforward way. In con-
trast, longitudinal models make it possible to introduce forward-looking elements to individ-
ual behaviour. Zaidi and Rake (2001) also draw a distinction between models where the rela-
tionships are entirely determined by the parameters defined within the model (deterministic
models) and models where random processes are also applied (stochastic models).
Most of the micro models under study in this project are dynamic, cross-sectional microsimu-
lation models with dynamic ageing (see the summary of the simulation properties in Table 2).
Exceptions are the Danish LAW and the stochastic lifecycle model of the Finnish Centre for
Pension, both auxiliary models from a pension modelling perspective, which provide cross-
section snapshots of new retirees benefit distributions. Other exceptions are static models, the
REDIS project of Luxemburg and the Belgian MIMOSIS, which are used for simulating and
comparing policy scenarios. The Dutch MICROS is capable of performing both static and
dynamic ageing. Below in a separate subsection of Section 3, we will return to the issue of
micro databases.
The initial databases used for these simulations may be various. While standard simulation
models are most frequently based on administrative data, microsimulation models tend to rely
on samples from censuses or surveys. Some tools use administrative databases or samples
thereof, others take surveys as a starting point. The crucial issues are whether surveys are rep-
resentative and whether one database (survey or administrative) contains the level of detail
necessary to capture all relevant attributes of individuals. The matching of databases provides
some help in both of these matters. Out of the 16 micro models covered in this report, 10 rely
on matched databases.
Dynamic models with dynamic ageing carry individuals through life-events. The level of de-
tail provided in the initial database determines to a great extent what life-events may be
tracked in the simulation (see the summary of life events tracked by administrative instru-
ments in Table 3). Individuals’ age and sex, as in standard models, is known of course, so all
models include birth and death. The Dutch MICROS, the Belgian MIDAS, the French
PRISME and DESTINIE simulate school-leaving age, first "marriage"/cohabitation, recou-
pling, rupture, and children born. With the exception of PRISME, they provide detailed in-
formation on household structure, including size and combined income. DESTINIE also
simulates kinship ties between non-core people, for instance the survival of parents or siblings
and applies differentiated mortality rates depending on age at leaving school. SESIM provides
the highest level of detail on the education front, while also tracking the age at which children
leave home. Hungary’s NYIKA model is only concerned with life and death among the
demographic parameters.

14
Table 3: Summary of life events tracked by selected administrative microsimulation models
Education Marital status Birth, adoption Labour market Occupation Retirement Health Other: …
position
AVID 1996, labour market states : 1) employed, subject to social insurance contributions, 2) child care, 3) housekeeping, 4) nursing relatives in need of care, 5) long-term sickness, 6) unemploy-
2002/2004 ment, 7) old-age pensioner, 8) unpaid work in family businesses, 9) low-paid employment, not subject to social insurance contributions, 10) civil servant, 11) self-employed, 12) lim-
ited earning capacity, 13) education, 14) other ; working hours and income position are also modelled for the relevant states, as well as the specific category in case of self-
employment
CAPP_DYN compulsory, upper single, mar- mortality, fertility, type of status (full dependent (private Old-age; disability, three disability
secondary school, ried/cohabiting, migration, exit from time, part time, out or public), self- survivor, social statuses based on
tertiary school (3 divorced, widower household of origin of the labour market, employed, assistance classification of the
year/5 year degree) unemployed), type ISTAT survey
of contract (typical,
atypical, permanent,
fixed term)
DESTINIE school leaving age first "marriage", all births from suc- Transitions between Transitions between Transition to retire- An ad hoc module Since the model
recoupling, rupture cessive unions (up to employment, unem- self- ment simulated on on disability had simulates complete
6) ployment and non- employed/private the basis of a behav- been added to Desti- kinship ties, we have
participation (includ- sector em- ioural model nie I. It will be information not only
ing pre-retirement) ployee/public sector probably updated for on what directly
employee, but only Destinie II. An happens to the indi-
in the new version of additional module is vidual but on what
the model (in version currently being built happens to related
1, sectoral allocation for the projection of individuals (e.g.
was considered fixed health expenditures death or disability of
all over the career) his parents...)
MICROS finish educa- divorce/marriage birth child(ren) become unem- finding a job/change early retirement / migration/housing
tion/leaving school ployed/disabled hours worked retirement market (rent/buy)
MIDAS ISCED levels 1-5/6 single, married, birth in work, disabled, legisla- wage-earners retire- reporting being
living in cohabita- unemployed, retired, tors/officials/manage ment scheme, civil chronically ill
tion, divorced, other inactive rs; professionals; servants retirement
widow tech. and ass. prof.; scheme, old-age
clerks; service and guaranteed minimum
sales workers; income
skilled agric. work-
ers; craft workers;
plant operators and
assemblers; elemen-
tary occ.
NYIKA school leaving age empirical probabili- empirical probabili- Transitions between No Transition to retire-
ties ties employment, unem- ment simulated on
ployment and non- the basis of past
participation patterns
PENSIM2 Yes Yes including co- Birth Yes Yes Yes disability and institu- mortality

15
 habitation but not tional care Pensioner savings,
same sex partner- housing, taxes and
ships benefits
PRISME school completion yes, including hou- children completion 1) labour market - working life comple- death completion
sehold structure and participation tion
kinship ties CNAVTS, 2) LMP
CNAVTS-
analogous, 3) LMP
non-CNAVTS, 4)
unemployment 5)
sick leave, 6) dis-
ability, 7) other
SESIM - upper secondary cohabitation, mar- adoption, fertility, unemployment, wealth accumula-
(drop-out, complet- riage, separation, children leaving employment, miscel- tion, housing
ing) divorce home laneous status
- university (drop-
out, completing)
- labour market (to
university or adult
education)
- adult education (to
university)

16
The sophistication of labour market positions is equally critical in these models. All models
simulate labour market entry, unemployment, disability and retirement. NYIKA distinguishes
6 labour market states depending on the frequency of being employed beside the above char-
acteristics. The list of labour market positions is highly detailed in AVID including 14 differ-
ent states (employed, subject to social insurance contributions; child care; housekeeping;
nursing relatives in need of care; long-term sickness; unemployment; old-age pensioner; un-
paid work in family businesses; low-paid employment, not subject to social insurance contri-
butions; civil servant; self-employed; limited earning capacity; education; other. The French
PRISME and the Hungarian NYIKA are both capable of handling employees participating in
different (public) pension schemes. In PENSIM2 of the United Kingdom the self-employed
are treated separately. MICROS, MIMOSIS, REDIS and PENSIM2 can also handle post-
retirement work, while in MICROS it is also possible to change the hours worked, which al-
lows incorporating part-time work.
As a general rule, transitions between states are simulated using estimated probabilities from
the initial dataset. Other, auxiliary datasets are often used to estimate certain transition pa-
rameters (e.g. LFS data in the case of PRISME and NYIKA, ISFol Plus survey on educational
attainments and work income in the case of Italy’s CAPP_DYN). Predicted probabilities
based on micro data are calibrated to long-term demographic projections and official labour
force forecast, as is the case of DESTINIE or PENSIM2. Looking beyond the aforementioned
deterministic method, economic and demographic transitions among states can be simulated
by Monte Carlo processes, like in CAPP_DYN and CeRPSIM. MIDAS, finally, applies
alignment to have its result being in line with exogenous retirement probabilities, in this case
pertaining to the AWG projections.
There are also various approaches to simulating the retirement process itself. The most basic
technique (and the most common one for that matter) is to rely on estimated probabilities as in
the case of other types of transitions. In SESIM, on the other hand, retirement is modelled
endogenously based on optimisation of pension wealth. CAPP_DYN also applies measures of
inter-temporal optimisation based on net social security wealth, but only for early pensions
(below statutory retirement age) and alongside an exogenously set adequacy condition (re-
placement rates need to reach a certain threshold). DESTINIE goes even further and offers
several behavioural patterns relating to retirement decisions: 1) retirement as soon as fulfilling
eligibility conditions for a full rate pension; 2) retirement based on utility maximization with
an income-leisure trade-off; 3) retirement based on relative changes in the pension wealth.
Most micro models deliver the results that standard models produce, such as the number of
retirees, contributors, aggregate indicators of revenues, expenditures and aggregate as well as
longitudinal individual replacement rates. On top of that, microsimulation models are capable
of producing indicators on intra-generational as well as intergenerational distributive effects
(in terms of replacement rate, or internal rate of returns). Income-related measures of poverty
are based on the Leyden poverty line, and various median poverty lines. Measures of inequal-
ity include the Gini, Theil, and decile ratios.
In models, which are founded on integrated micro databases, comparisons between labour and
social security income (possibly also other types of incomes) is also feasible, while some even
have the potential to move from the individual level to the household and take into account
alternative sources of income after retirement, including private savings (AVID, LAW). Other
models (PRISME is an example) are able to track the effects of reductions in the case of early
exit and supplements for working longer years.

17
3. Approaches to pension modelling

In this section we characterise the approaches of administrations to pension modelling. In

contrast to Section 2, the unit of analysis is the administration, not the model. The emphasis is
not on model properties but on the way the process of modelling is embedded in the adminis-
trative work. We will use the simple taxonomy of Table 4.
The crucial dimension of the classification is the absence or presence of instruments process-
ing micro datasets. We found 16 administrations that do not apply microsimulation models in
pension modelling. This is not to say that they do not use micro data or microsimulations at
all; it may well happen that microsimulations are available for the administration in these
countries for other purposes. We discuss them in Section 3.1, whereas countries applying mi-
crosimulation models in pension projections are dealt with in Section 3.3. Section 3.2 gives a
summary of the micro datasets in use.
The other dimension of the taxonomy is loosely related to the modelling capacity of the ad-
ministration. Countries that use only standard models may import available general-purpose
instruments or they may develop their own. Modelling capacity also affects administrations
applying microsimulation. Some of them outsource this activity, while others do it by them-
selves or in cases they organise special units for such purposes and make microsimulation
their primary modelling tools.

Table 4. Classification of administrations according to modelling pensions

Exclusive use of standard models
Imported models customised to special needs Bulgaria, Cyprus, Latvia, Lithuania, Malta, Portugal,
Slovakia (MoF)
Own developments
No steps toward microsimulation Estonia, Greece, Ireland, Poland, Slovakia (MoLSAF)
Slovenia
Steps toward microsimulation Czech Republic, Spain, Hungary, Austria
Utilisation of microsimulation models
Microsimulation models as complements
Modelling work contracted out Germany, Italy
Modelling work by administration Belgium, Denmark, Luxemburg, Netherlands, Finland
Microsimulation as primary modelling tool France, Sweden, United Kingdom
Note: No information was available on Romania.

3.1 Exclusive use of standard models

Many administrations do not use micro data directly in pension projections but rather base
calculations on cohort averages, fictitious typical careers or macro models. The models ap-
plied in these cases are either imported from international agencies, such as the International
Labour Organisation (ILO) and the World Bank (WB), or developed by the administration
itself. We will handle these two groups separately in subsections 3.1.1 and 3.1.2, respectively.
In subsection 3.1.3 we collect some experiences of administrations that have made the neces-
sary steps of introducing microsimulation techniques in pension projections.

18
3.1.1 Imported models customised to special needs
Of the EU national administrations, Bulgaria and Cyprus reported to have applied the ILO
pension model. Bulgaria employs the entire ILO Social Budget Model, of which ILO PENS is
a sub-program. It is employed in the annual budget planning but the administration also tests
policy reforms in terms of the budgetary consequences and makes medium-term projections.
In Cyprus the model is updated tri-annually and employed in actuarial valuations and long-
term budget planning. Another frequently used model is PROST (Pension Reform Options
Simulation Toolkit), a general-purpose model of the World Bank, which has been applied in
over 80 countries. This number includes nearly all EU Member States with the exceptions of
Bulgaria, Latvia, the Czech Republic and the United Kingdom. Responding to our question-
naire, however, only two administrations, Malta and Slovakia, mentioned this toolkit as cur-
rently applied to pension modelling. In Malta, it is not part of the routine budgeting process
but mainly used in long-term planning and estimating the effects of various policy scenarios.
In Slovakia, a background institute of the Ministry of Finance employs PROST, whereas the
Ministry of Labour, Social Affairs and Family has its own model, MAJA, which was devel-
oped in cooperation with the ILO and the World Bank. A further development by the World
Bank is PRISM (Pension Reform Illustration and Simulation Model). It has been in use in
various Central-Asian countries and in Eastern Europe (Czech Republic, Hungary, Mace-
donia). Among EU countries currently Lithuania applies PRISM. Portugal also reported on an
imported model, an adaptation of ModPens, developed by the Spanish FEDEA. The Portu-
guese version is updated on an annual basis and used for the assessment of financial sustain-
ability of the social security system. The Latvian administration might also use an imported
tool from Sweden. The PENMICRO team got access to a presentation but no answer was re-
ceived from the administration as to this aspect of the model.
Countries, which imported macro or cohort-based models, can be further divided into two
subgroups, depending on how sophisticated their projection methodologies are. In the first
subgroup there are countries where the modelling instrument does not differentiate between
stock and entry pensions but it simply extrapolates the number of stock beneficiaries, usually
assuming there to be a constant proportion of the cohort population at different ages. Such
models, like the ModPensPor, can take on board neither the effects of changing labour market
situation nor the implications of regulatory amendments; hence they are often complemented
by external adjustments (’expert judgement’). These countries may want first to consider the
advantages of introducing a distinction between stock and entry benefits. This would enable
them to capture certain effects of legislative moves or labour market developments, which
gradually filter into stock pensions through changes in the number and average amount of
entry benefits. Data on new benefit awards broken down by gender and the age of retirement
are available in all EU Member States, although in certain cases they can only be collected
from more than one agency. Obviously, the modelling framework should also be adjusted so
that it can handle entry benefits separately from stock benefits.
Other countries use more developed imported models that already include this distinction be-
tween entry and stock benefits, such as the World Bank’s PROST. These cohort-based models
usually calculate entry pensions based on average characteristics as observed in the past.
However, in most cases they assume these characteristics (e.g. length of contributory period,
average wage assessment base, household composition etc.) to remain constant throughout the
projection horizon. These models are usually apt for simulating various effects of certain pa-
rametric changes (for example, a change in the method of benefit indexation) and, although to
a very limited extent, might be able to take account of recent trends in employment patterns.
Still, these instruments have major shortcomings when it comes to the simulation of legisla-
tive modifications in the field of benefit calculation, or there are trend changes in the labour

19
market or significant developments with regard to scheme coverage. Most of these models fail
to take on board the implications of past labour market trends and the resulting shifts in pen-
sion entitlements. Even if some aggregate data on contribution history are available at the
pension or tax authorities, these models usually cannot handle such input. Different character-
istics of migrants or changes in educational attainment are as a rule omitted by these instru-
ments, even on an aggregate level.
In return for these shortcomings, imported models are a useful first step in pension modelling.
They are easy to adopt by the administration and the costs involved are also modest. How-
ever, for the same reasons their applicability is limited. Since they are usually made for gen-
eral use, they frequently resist more specific analysis. In particular, they are usually difficult
to parameterise with regard to interaction of various effects, such as the effect of education or
family structure on labour supply and wages. Consequently, these models are frequently no
more than simple projections of the current situation combined with ad-hoc parameterisation
of the future. In addition, adaptability often entails that imported models cannot accommodate
all subtleties of the national legislation. These are not problems in themselves as long as in-
terpretation of the results is constrained by these limits.
Furthermore, adopting a model implies a reliable knowledge by the administration of how to
use it, but does not necessarily create capacity for further development or modelling for more
specific purposes. As the model was not developed from scratch by those who become re-
sponsible for taking care of it, the risk remains that the tool is regarded as a black box. This
may increase the likelihood of misinterpretation of results, especially in circumstances when
the impacts of regulatory changes are assessed. It is crucial in such cases to build up a model-
ling capacity by the administration of its own as well as to create a routine, which assures that
the model outcomes are used in long term planning.
For all these reasons, countries using imported models may want to consider the construction
of an instrument of their own development. This does not mean that imported models ought to
be discarded at once: work can be started with rewriting a certain module, if the general
framework still seems to be appropriate. Below we list some tools developed by other coun-
tries that might offer useful hints about projection techniques. Areas where further develop-
ment would generally result in significant improvements include the following three points.
Aggregate administrative data on past pension accruals allow for valuable conclusions on the
likely evolution of new benefit awards. Historical data on the acquisition of pension rights
could also help in establishing a tentative link between employment (contribution payment) or
participation rates on one hand and arising pension rights on the other hand. The effects of
certain policy measures - such as the gradual introduction of notional defined contribution
schemes - cannot really be simulated without data on past accruals. A number of examples
can be mentioned on how cohort-based models can incorporate information on pension rights
acquired in the past. The Danish pension model is based on data gained from administrative
panel data. The Italian RGS model also makes use of historical data about workers with a con-
tribution past.
Countries, which are not in a position to collect real career profiles, may also try to feed their
cohort-based instruments from typical agent simulations. In certain cases this may also lead to
a better estimation of trends on the average amount of newly granted benefits. For example,
when a final salary scheme is gradually turned into an average salary scheme (i.e. the wage
assessment period of the pension calculation is extended), new pensions will grow at a lower-
than-earlier rate due to the normally concave shape of the longitudinal wage path. The magni-
tude of such effects can also be estimated with ’typical’ theoretical careers and then the results
imputed to the cohort model. For example, a similar approach was used in Hungary in the

20
model operated by the Ministry of Finance. In the Belgian MALTESE the average amount of
newly granted old-age pension was also estimated for a number of careers, which were con-
sidered typical.
Cohort-based models as such distinguish between groups within the population based on age
and gender. These groups are usually further divided along the dimensions of employment
status, benefit type and sometimes according to other aspects, as well. However, important
characteristics of these sub-groups are not always taken into account. For example, most co-
hort-based models do no apply mortality rates differentiated across various statuses, although
mortality rates can differ significantly according to health status, ethnic origin, marital status
etc. Some models have already applied such differentiation with success. For example, the
Finnish Centre for Pensions used a mortality projection adjusted for disabled people. Simi-
larly, the pension model of the Ministry of Finance of Hungary made a downward adjustment
to mortality rates of disability pension beneficiaries and an upward adjustment to all other
cohort members as compared to cohort average rates, based on historical averages. On the
contribution side, the Danish pension model imputes different earnings levels and employ-
ment rates to immigrants or descendants of immigrants. The Lithuanian version of the PRISM
model distinguishes 5 earnings categories (however, we could not clarify whether mortality
rates have also been differentiated according to these categories or the benefit amount only).

3.1.2 Own developments

Many administrations applying only standard models have developed instruments specifically
for their own purposes. Some governments belonging to this group have signalled steps to-
ward microsimulation. These are the Czech Republic, Spain, Hungary, and Austria. Those
that do not belong to this sub-group may also have started working in this direction but they
have not informed us. The countries belonging to this group are Estonia, Greece, Ireland, Po-
land, Slovakia (Ministry of Labour, Social Affairs and Families, MoLSAF) and Slovenia.
Own developments vary by type, sophistication and the way models are administered. Poland
and Greece assign the task to institutes specialised in pensions, such as the Social Insurance
Institute and the National Actuarial Authority, respectively. Consequently, these models are
more focused on pensions with no special reference to the rest of the tax-transfer system. In
Poland, even the farmers’ pension scheme is set aside and projected by Polmodel, a simula-
tion model of the social policy budget, which is not under the authority of the social insurance
system. In Greece, the need for modelling replacement rates appeared only in the framework
of the ISG exercise. The resulting model calculates only theoretical replacement rates and its
scope is limited to IKA-ETAM, the largest social insurance fund, which, as mentioned above,
covers about 55 percent of the employed population.
The second group comprises models, which are operated by government ministries. This is
usually the Finance Ministry (Estonia, Hungary, Ireland, Slovenia and Spain), or the Ministry
of Labour and Welfare (Austria and the Czech Republic). The Slovakian administration ap-
plies two models one by the Finance Ministry (see above) and one by the MoLSAF. The latter
is an actuarial model with little reference to current or future labour market trends or individ-
ual characteristics other than age and gender. The Czech pension model differentiates, beyond
age and gender, between employment types (employed/self-employed). In addition, old-age
benefit projections are also disaggregated by income deciles. With regard to entry pensions,
cohort members are classified by earnings and career lengths although these parameters are
not realigned with prospective changes in the projected labour market situation. The Estonian
model is by one degree even more complex in that it also incorporates an age and gender-
specific relative wage structure. The internal consistency of the parameters is assured by a

21
partial equilibrium model. In contrast, the Slovenian SIOLG is a complex dynamic overlap-
ping generations (OLG) model with a general focus on sustainability of the entire tax-transfer
system, even though intergenerational transfer schemes, such as pensions, health care and
long-term care, can be analysed separately. This assures a more consistent parameterisation.
Of the countries that reported on steps made toward microsimulation, we learned that the
Czech administration is currently in the process of preparation of the databases that would
allow enriching the analytical capabilities of the government and building a microsimulation
model complementing or replacing the currently available standard cohort model.
In Spain the Social Security Administration disclosed the records of 1.2 million people, 4 per-
cent of the reference population, containing information on individual work histories (em-
ployment, wages, contributions) and pensions. This dataset can be extended with additional
fiscal and census information by matching it with the population registry and the income tax
databases. We received conflicting information as to whether this matching has actually oc-
curred or just planned.
In Austria a microsimulation model is already under construction by an external contractor of
the Federal Ministry of Social Affairs and Consumer Protection. It will be based on various
micro datasets such as the data on the pension stock (Stand Pensionen, SP), the annual statis-
tics of pensioners (Pensionversicherung Jahresstatistik, PJ) and the condensed insurance his-
tory of pensioners (Verdichteter Versicherungsverlauf Pensionen, VVP). The SP set covers
the whole population and contains data on pensions and various other allowances such as e.g.
childcare. The PJ dataset, also covering the entire population, contains various pieces of in-
formation required for calculating entry pensions as well as exits from the system due to e.g.
loss of disability status or death. The VVP set, in addition to information derived from previ-
ous data, also contains the entire employment and wage career of the people covered. All
three datasets include social security ID numbers, which lend the opportunity to directly
match the data with tax records and other administrative data. Here again, we are uncertain if
this matching is just a possibility or it has actually happened.
In Hungary the set of models available for the administration have been recently extended
with the NYIKA model. Whereas the Ministry of Finance and the National Bank apply stan-
dard cohort models, the NYIKA is based on a large administrative panel dataset, KELEN,
containing information on contributors to the pay-as-you-go pillar. The data cover the period
of 1997-2006 and include over 6 million records (about 4.2 million each year). It is matched
with contributory records of the mandatory private funds of the second pillar (KPN). The
model also gains further information on current pensioners (from the NYUFUR micro dataset
of the pension administration) and contributors (from the Labour Force Survey). The first
results produced by the NYIKA have been recently released. Despite the current focus on
assessing the extent of future old age poverty, the model can also be used for the analysis of
fiscal sustainability. It belongs to the family of imported models in that modelling was con-
tracted out to an international consultancy. The administration currently does not have the
capacity neither to maintain nor to further develop the model. This and the data background
(since 2007 individual contribution files are no longer held by the pension administration)
make the future of the model uncertain. A further weakness is the actuarial nature of the
model.
Cohort models developed in-house are usually considered to provide a fair estimation on the
aggregate number of beneficiaries, revenues and expenditures, while they are still able to pre-
serve some degree of transparency and relative ease in projection techniques. However, the
cohort-based methodology as such has but limited means for capturing effects of various dis-
tributional changes. This comes as a drawback not only when various indicators of income

22
inequality are being looked for, but often leaves the main outcomes (revenues, expenditures)
of the modelling distorted. Let us highlight some issues that would pose a challenge to most
cohort-based instruments.
Countries that experience marked changes in the development of migration will be faced with
difficulties in producing reliable budgetary projections with simple standard models. For ex-
ample, entitlements for state pensions based on the length of residence (where such schemes
exist) depend heavily on the age structure of new migrants. Furthermore, immigrants tend to
have participation rates and educational attainment different from the indigenous average in
most countries, which might have a sizeable impact on the earnings related components of
social security benefits.
Second, changing patterns of behaviour in the field of marriage, cohabitation and divorce can
also have important implications on pension spending and income in old age. In a couple of
countries, single persons may receive lower or higher benefits than those with a spouse or
partner. If the number of single persons among the pensioner population rises, the cost of pen-
sions might also go down or up depending on the benefit formula. Survivors’ benefits will
also be affected by changing household structures. In a number of countries, the ever more
prevalent partnership forms are still treated more or less differently when it comes to pension
entitlements, as compared with marriage.
Third, the composition of society based on educational attainment is also expected to bring
about changes in the pensions landscape. In most countries, people with only basic education
can count on a lower-than-average life expectancy and a significantly higher probability of
getting disabled before reaching standard retirement age. Hence, the growing share of people
with higher education is likely to alter take-up ratios and average beneficiary characteristics in
the future.
Fourth, changes in the labour market (both past and future) will have a substantial impact on
the evolution of newly granted pension benefits. Especially countries that have experienced
major economic changes in the last decades will see changes in the composition of their popu-
lation according to acquired pension entitlements. For example, in the majority of new Mem-
ber States in Central and Eastern Europe the labour market was hit severely in the 1990s fol-
lowing the collapse of the communist era. Cohorts retiring today started their careers some-
time around 1965-1970, so they have spent about half of their career prior to the regime
change and another half in periods characterized by far lower employment rates. The distribu-
tion of people within subsequent cohorts according to the length of contribution period varies
quite significantly year by year since the fraction of working lives spent in the pre-1990 years
is shrinking as we proceed in time. The assumption of constant entry characteristics would
therefore mean a simplification with adverse effects on projection results.
Cohort-based models are awkward in taking on board the effects of such socio-economic
changes. Even if modellers have access to data on retrospective pension accruals, the process
of pension rights acquisition can only be simulated in a limited form if only a standard model
is available. For example, the link between the evolution of the labour force, which is nor-
mally imputed to standard models from an external source, and the acquisition of pension
rights is far from being evident. A one per cent increase in the age-specific employment rate
from one year to another does not give way to straightforward conclusions about the devel-
opment of pension rights within that cohort. Do those already employed in the previous year
work more hours or more frequently or do other people enter the market who had been unem-
ployed or inactive beforehand? The answer would be vital in assessing cohort-average pen-
sion accruals at retirement, not to mention the distribution of accrued rights within cohorts,
which would also be essential for giving an appropriate estimate of overall expenses. For ex-

23
ample, in countries where there exist minimum requirements on the lengths of service or con-
tributory period, the share of those fulfilling these requirements can ideally be approximated
on the basis of such distributions. The same applies to the modelling of taxation: without in-
formation on intra-cohort distributions, projections on average tax rates may easily become
distorted. For the same reason, models built on cohort averages are incapable of giving a good
estimation on how the share of people eligible for a top-up to minimum benefit amount will
evolve in the future.
Therefore, administrations working exclusively with standard models can be suggested to
make a step further and turn their attention to microsimulation models, which have better ca-
pabilities in tackling such distributional effects. The first step towards this would be taking
stock of existing administrative databases and surveys to find out what efforts would be
needed to use them for the setting up of new microsimulation instruments. While the software
background for microsimulation models can be imported from abroad as well as for cohort-
based tools (if necessary), good quality data on all individuals or at least on a sufficiently
large sample of individuals is essential.

3.1.3 Experiences of introducing microsimulation to the modelling toolkit

The range and quality of micro level data available for the exercise determines the complexity
of the model that can be constructed, along with a number of other factors, such as the allotted
timeframe, financial resources, personnel, etc. Evidently, the objective of the model should be
set carefully and one should avoid being overly ambitious in situations of data constraints or
limited resources. Setting priorities is key in starting with a new development. For a start,
modellers may want to consider focusing on selected fields instead of aiming from the outset
at a comprehensive model. The most common simplifications in practice are the following.
Initially, modelling work could be limited to certain elements of the pension system, with the
aim to extend it with new schemes or components later on. For example, the SADNAP model
(Netherlands) is currently capable of projecting state pension expenditures, but there are plans
to upgrade it later to cover private pensions, as well.
Most microsimulation models take the form of a relatively flexible modular structure, which
allows for a more gradual development. At the first stage, models can be regarded as opera-
tional even if certain modules would need further refinement but otherwise the instrument is
able to produce some basic analyses it is primarily aimed at. Developers should, however,
take care of preserving the capacity of the model for later upgrades and extensions. The
modular structure also increases the transparency of the model. For example, the SADNAP
model currently uses a simple random procedure to distinguish between single pensioners and
couples, but the construction of a proper marriage market module is already envisaged. In the
case of the Belgian MIMOSIS, an elaborated module for calculating pensions based on past
career information is currently being developed and shall be integrated into the next version.
A similar extension is planned in the framework of the REDIS project of Luxemburg.
Beyond these initial or temporary simplifications and building strategies, the setting up of a
new microsimulation model can also be assisted by professional help from those who have
already had experience with microsimulations. This can take the form of a personal collabora-
tion, a referral to literature on modelling experiences or – most directly – the usage of plat-
form or source code of an already existing foreign or academic microsimulation instrument.
This latter option can considerably accelerate the process of model construction. For example,
SESIM (Sweden) utilized the CORSIM (US) methodology first developed at the Cornell Uni-
versity. In Hungary, the NYIKA microsimulation is built on the Prophet actuarial software,

24
which has been in use in several countries for the projection of defined benefit occupational
schemes. The REDIS project took the static EUROMOD simulation as its starting point for
their plans to develop a dynamic microsimulation model with enhanced capabilities in the
area of social policy analysis.
Countries might also want to combine their efforts in developing a good quality dynamic mi-
crosimulation. Currently, this is the case with the MIDAS project, encompassing Belgium,
Germany and Italy. The instrument also has the potential to be utilised in an international con-
text and produce comparable outputs. While datasets need to be supplied from national
sources, the main processes can be programmed in a joint effort, thus sharing costs and exper-
tise. Collaboration between governments and academics can also boost the chances of a new
microsimulation tool. For example, SESIM as well as the PENSIM2 were supported by re-
nowned academic experts.

3.2 Micro datasets

Social security administrative databases most often contain information on employment /
wage / contribution history, which could be a good starting point for building up a micro data-
set (see the various examples of micro data utilisation summed up in Table 5). The Swedish
MiMESIS, the oldest of the microsimulation models sampled here, and the Hungarian NY-
IKA, the latest development, both use only administrative data of the social insurance system.
In the Hungarian case it means matching two sets, one of the administration of the pay-as-
you-go scheme (KELEN) and another one of the supervisory agency of mandatory private
funds (KPN). In subsection 3.1.2 we have given a brief description of the KELEN; further
details can be found in Section 4, the model inventory.

Table 5:
Examples for utilisation of micro datasets in administrative microsimulation models
social insurance administrative data
Hungary-NYIKA (KELEN and KPN)
Sweden-MiMESIS (ATP)
social insurance administrative data matched with tax records, census data
Spain
Netherlands-SADNAP (Statistics Netherlands and Social Insurance Bank)
Austria (SP, PJ, VVP)
Finland-Finnish Stochastic Life-cycle Model (Statistics Finland and Finnish Centre for Pensions)
large household survey
France-DESTINIE (INSEE Financial Assets Survey (Enquete Patrimoine))
Netherlands-MICROS
social insurance administrative data matched with household surveys
France-PRISME (CNAV and LFS, the Inter-scheme Pensioners Sample, Family History Survey)
Luxemburg-REDIS (Luxemburg Social Security Data Warehouse and EU-SILC)
United Kingdom-PENSIM2 (LLMDB, BHPS, FRS)
matching numerous administrative datasets
Belgium-MIMOSIS (Data Warehouse of the Crossroads Bank for Social Security)
Denmark-LAW (Statistics Denmark)
Sweden-SESIM (LINDA)

However, the analysis based solely on social insurance data often encounters difficulties. The
scope of such databases is usually limited. They allow projections of future contributors,
beneficiaries, revenues and expenditures but do not give the chance for simulating the effects
of changing enrolment rates, household formation and even certain labour market develop-

25
ments. In addition, projected income distributions and calculations of poverty measures are
limited, since alternative income sources as well as household and kinship resources are not
accounted for.
Some of these details can be gained by matching social insurance data with other administra-
tive datasets, such as tax records or census data. If legal constraints do not hinder it, this can
be carried out relatively easily either through natural identification numbers or by the use of
artificial connection codes. For example, the Finnish stochastic life-cycle model links two
separate administrative databases (pensions related information from the Finnish Centre for
Pensions and employment data from Statistics Finland). As mentioned above in Section 3.1.2,
this is the road taken by Spain and Austria as well. The Dutch SADNAP model is also based
on contribution data from Statistics Netherlands and benefit data from the Social Insurance
Bank (SVB). Further details of the data background of this and selected other models can be
found in Table 6.
Census data extend the information set available for analysis but they still often lack detail,
which could only be gained from surveys. The Dutch MICROS is based on a dataset consist-
ing of more than 60,000 households. It comprises information on household composition,
income and housing costs. The French DESTINIE has its background in the Enquête Patri-
moine, a household survey covering 65,000 people. Such large survey samples are rich in
information but if the analysis is multidimensional, cell frequencies can be too small and
standard errors too high. In addition, such surveys need re-weighting to macro controls not
only in the calibration phase of the simulation, but also in the base year. The Enquête Patri-
moine, for instance, is adjusted to survey and administrative data such as the Labour Force
Survey, the Family History Survey, the FQP (survey on training and skills), the Annual Dec-
larations of Social Data, the Permanent Demographic Sample and the Fiscal Income Survey
(see Table 6).
A further step ahead is matching social security administrative data with surveys. This is the
path taken by the Model Development Unit of the Department of Work and Pensions of the
UK government. Their model, the PENSIM2 is based on three data sources. The Lifetime
Labour Market Database (LLMDB) is a panel of 1 percent of the individual population (1975-
2006) containing information on state pension contributory histories, private pension mem-
bership (some schemes only), employment, and wages. The British Household Panel Study
(BHPS) is a longitudinal household survey based on annual interview 1991-2006 and the an-
nual cross-sectional Family Resources Survey (FRS) containing data on housing, demogra-
phy, household composition and other contextual information. The three data sources are
matched by a statistical procedure, not by personal ID numbers, to form a dataset of synthetic
records.
The Luxemburg administration also matched administrative data drawn from the Luxemburg
Social Security Data Warehouse with the Socio-Economic Panel “Living in Luxem-
burg”/European Union Statistics on Income and Living Conditions (PSELL3/EU-SILC) sur-
vey. The administrative data contain information on the whole population of Luxemburg
(449,000 observations for residents in 2003). It is comparable to the data warehouse of the
Belgian Crossroads Bank (see below), although it is simpler. It contains no information on
education, wealth and capital income but it will be extended with data on health care and
long-term care utilisation in the future. The survey data is derived from a representative sam-
ple of around 3,600 private households (9,800 individuals) residing in Luxemburg. It com-
prises detailed information on income, household structure and socio-economic information.

26
Table 6: Micro datasets processed by administrative microsimulation models
model country name of the dataset administrative units in dataset sample size (as level of infor- sampling scheme adjustments work and earn-
or survey % of popula- mation (elimination of ings history
tion) bias, alignment)
AK-analysis Finland AK administrative appr. 5,000 individu- appr. 0.1% individual 400 to 1000 people detailed career
als from every fifth birth- information, in-
cohorts between 1905 cluding earnings
and 1970 (1963-2005)
AMM Austria SP (pension stock), annual administrative population total population individual condensed insur-
statistics of pensioners (PJ), ance history
insurance history (VVP)
AVID, 1996, Germany individual pension records administrative net sample size appr. appr. 0.02% individual + 1996: sample of people entire work history
2002/2004 and survey 14,000 people spouse having a social security of sampled indi-
record and eligible for viduals
old-age pension at 65
representing cohorts
born between 1936-
1955 surveyed;
2002/2004: survey data
of sample representing
cohorots born 1942-
1961 matched with
social security records
CAPP_DYN Italy SHIW_02 (Bank of Italy's survey 21,148 individuals appr. 0.04% individual two-staged: in the first post-stratification working path and
Survey of Households Income within 8,011 house- stage, municipalities procedure applied: earnings are recon-
and Wealth, 2002) holds selected in 51 strata, in original sample structed retrospec-
the second step, house- weights adjusted on tively by means of
holds selected randomly the basis of the last econometric esti-
within each stratum ISTAT census mations
DESTINIE France Financial Assets Survey (En- survey 65,000 individuals appr. 0.1% individual + calibrated to survey
quête Patrimoine) household + and administrative
kinship data such as the
LFS, the Family
History Survey, the
FQP (survey on
training and skills),
the Annual Decla-
rations of Social
Data, the Perma-
nent Demographic
Sample and the
Fiscal Income
Survey
MIDAS_BE Belgium Panel Dataset of Belgian survey 8,488 individuals appr. 0.08% individual + na. na. complete history of
Households household earnings (1994-
2002) for the
selected sample of
individuals

27
MIMOSIS Belgium Datawarehouse Labour Market administrative 100,000 individuals appr. 3% individual random sample based na. data available on
and Social Protection completed with all on the National Register wages between
household members 1954 and 2001, but
no data on past
pension accruals
NYIKA Hungary KELEN (database of the Pen- administrative 6 million records total contributing individual retrospective data
sion Insurance of entitlements) population on contribution
and KPN (database of pension payment, 1996-
fund supervisory agency) 2007
PENSIM2 United LLMDB2 (Lifetime Labour administrative + 11,000 and 57,000 1% individual + National Insurance British Household 1975/1978
Kingdom Market Database) matched with survey samples of LLMDB2 household Recording System Panel Study
the Family Resources Survey, a used for simulation. (NIRS) containing (BHPS) longitudi-
standard cross-sectional house- details of National nal survey data as a
hold survey Insurance records for “bridge” to im-
over 60 million indi- prove the match of
viduals LLMDB2 and FRS
PRISME France CNAV administrative 4.4. million records 5% of registered individual
(5% sample); first population
results published on between 1900-
1% sample) 2005
REDIS Luxemburg Luxembourg Social Security administrative + SSDW: 449,000 SSDW: 100% individual + random sample na. cumulated periods
Data Warehouse; PSELL/EU- survey observations; (whole popula- household of insurance; data
SILC survey PSELL/EU-SILC: tion); on wages reaching
9,800 individuals PSELL/EU- back to 40 years is
within 3,600 house- SILC: appr. 2.2% planned for the
holds dynamic MSM arm
of the project
SADNAP Nether- CBS micro data file (Statistics administrative CBS: 16 million na. individual random sample CBS data from information on the
lands Netherlands) on general old- records (whole popu- 2004 were simu- acquired years of
age pension (AOW) entitle- lation); SVB: 2.6 lated to 2006, entitlements is
ments; SVB data on existing million records (all which is the year of contained in the
beneficiaries (Social Insurance AOW beneficiaries) origin for the SVB CBS dataset
Bank) data
SESIM Sweden LINDA, based on Income administrative 380, 000 people appr. 3.5% individual + two distinct samples: data on pensionable
Registers and Population and (1999); combined household random sample of entire income available
Housing Censuses and ex- with household population and a no from 1960, on
tended with 9 smaller adminis- members: 786,000, overlapping 20% pension income
trative datasets sample of immigrants from 1974

28
The French PRISME model is built on a similar matched dataset as well. The sample base
contains 1/20 of the insured from generations between 1900 and 2005 (around 45,000 persons
by generation), and since 2006, the projections have been made on the 1/20 sample. The first
results published are based on a 1/100 sample. Data of the Labour Force Survey, the Inter-
scheme Pensioners Sample and the Family History Survey are matched to arrive at the com-
plete dataset.
The final step in this gradual improvement of micro datasets is matching social insurance ad-
ministrative data with other administrative datasets that contain information on subjects usu-
ally covered only by surveys, such as family composition, alternative social assistance reve-
nues, income, wealth or health. The Statistics Denmark created a panel dataset by matching
25-30 different administrative databases by the identification number of Danish residents.
Although these sets have the individual as their unit, address and some other features allow
the reconstruction of households.
The Belgian Crossroads Bank for Social Security maintains a data warehouse, which is cre-
ated by matching administrative datasets generated by the national bureaus for social security,
employment, health insurance, child allowances, industrial accidents and occupational dis-
eases, old-age pension and the national register as well as the regional employment services
and the public service for social integration. The data warehouse contains individual level
information on date of birth, gender, nationality, wages, employer, social security contribu-
tions, labour regime (full time / part time), unemployment spells, old-age pensions (benefits,
entry date), incapacities to work and child care allowance, among other pieces of information.
Household members can be linked. This background facilitates a sampling procedure of ran-
domly selecting a sizeable sample from the national register (100,000 people in the case of the
MIMOSIS model), extending it with household members and filling the dataset with the
available information.
A similar process can be executed in Sweden, where the LINDA (Longitudinal Individual
Data for Sweden) set contains information from 11 administrative registers properly matched.
The core of LINDA is the Income Register (tax files available retrospectively from 1968) and
the Population and Housing Census (repeated in every five years between 1960 and 1990).
These two files were extended, partly from the late 1970s but mainly from the 1990s, with
separate registers of pensionable income, pensions, unemployment spells and unemployment
benefit, higher education and adult education, sick-leave, parental leave and wages. We were
informed that the set is being broadened with health data as well.

3.3 Utilisation of microsimulation models

With reference to Table 4 in the introduction of Section 3 we turn to another group of Mem-
ber States, notably Belgium, Denmark, Germany, France, Italy, Luxemburg, the Netherlands,
Sweden, Finland and the United Kingdom, which made more advances in the use of mi-
crosimulation techniques in pension modelling. We distinguished two sub-groups within this
class: administrations that use microsimulation as a complement of their, sometimes highly
sophisticated, standard models and administrations that apply microsimulation as the primary
modelling tool. The borderline is not clear-cut and some countries could be put to either class.

Microsimulation models as complements

The general pattern in the first subgroup is to have a standard cohort model for budget plan-
ning and a microsimulation model for special purposes, such as evaluating reform scenarios,

29
calculating longitudinal replacement rates and measuring income distribution, inequalities and
poverty in old age. In some cases microsimulation is outsourced, which possibly reflects the
lack of capacities in the administration. Our impression was that Germany and Italy belong to
this subgroup. In some other cases, however, giving microsimulation a supporting role is an
approach, apparently. The administration may well trust the standard model as a reliable
source of budget forecasts and feels no urgency to replace it with microsimulation. We found
Belgium, Denmark, the Netherlands, Finland and possibly Luxemburg fitting in this sub-
group.
Germany employs an administrative expert team for regular projections of future revenues
and expenditures representing the relevant government agencies. However, in order to extend
the scope and go beyond aggregate cash flows and also to identify risk groups and simulate
the impact of various reform measures on the lower edge of the income distribution, it
launched the AVID project. AVID is a non-regular extension of the usual projection proce-
dure. So far it was conducted twice, in 1996 and in 2002/2004. The project aimed at selecting
a sizable sample of people having social insurance record and having been eligible for old-age
pension upon reaching the age 65 years. The sampling process set off with social insurance
records in the 1996 round, which served as the sample frame for the survey. In the second
round, in 2002, the order was turned to the opposite. A sample representing 20 cohorts of con-
tributors born between 1942-1961 was surveyed and this information set was extended with
social security records. The surveys extended the detailed information on employment career
held by the social insurance agency with a rich set of data on the household and alternative
income sources. The information collected were utilised in a dynamic microsimulation exer-
cise, which the administration, notably the Federal Ministry of Labour and Social Affairs and
the Federation of German Statutory Pension Insurance Institutes outsourced to small private
contractors. Apparently, this practice was repeated in the second round as well so the gov-
ernment have not internalised the simulating capacity yet.
The Italian administration, as we were informed, also extends the scope of its pension projec-
tions based on its own standard cohort model with microsimulations carried out by external
actors. In this case external actors are academic institutes, the Center for the Analysis of Pub-
lic Policy, a joint venture of the Universities of Modena and Bologna, which built up
CAPP_DYN, and the Centre for Research on Pensions and Welfare (CeRP) in Turin, devel-
oper of a family of models, including a microsimulation. Both are recent developments and it
was unclear to the PENMICRO research team of how these models, named by our Italian con-
tact, are actually employed in the administrative work. Both models seem to face a problem of
access to micro data. Both use relatively small income surveys, the Survey of Household In-
come and Wealth of the Bank of Italy (about 21,000 respondents) and the Italian EU-SILC,
which cannot be matched with administrative data due partly to legal constraints on data
matching and partly to difficulties to obtain such data to start with.
This is in sharp contrast with the situation in Belgium, Denmark, Luxemburg, the Netherlands
and Finland. In particular in Belgium and Denmark, authorities created large data warehouses
by matching various administrative datasets (see Section 3.2). In addition, these administra-
tions are well prepared for developing and maintaining microsimulation models of their own.
In Belgium, special government units, the Federal Planning Bureau (FPB) and the social secu-
rity administration (Federal Public Service Social Security, FPSSS) are responsible for such
exercises. The FPB has a range of models including a standard cohort model (MALTESE), a
standard typical agent model (MEP) and a dynamic microsimulation model (MIDAS_BE),
whereas the FPSSS uses MIMOSIS, a static microsimulation. MIDAS_BE works with the
Belgian PSBH, the production dataset of the ECHP; MIMOSIS is fed with a large sample of
matched administrative data.

30
In Denmark, the Ministry of Finance can also choose, depending on the research question,
which model to use. The ministry has a standard cohort-based actuarial valuation model and a
static microsimulation model. The former is drawn on when it comes to long term stability
issues, whereas the latter is consulted with if income distribution is the question. The opinion
of the ministry is that, in light of the institutional structure of the Danish pension system, this
variety of modelling instruments is sufficient. Access to data is less of a problem. The admini-
stration has an administrative dataset, created by matching 25-30 various smaller sets, of its
avail.
The Dutch administration has a variety of instruments to use for various analyses. GAMMA
(General Accounting Model with Maximizing Agents) is a dynamic overlapping generations
general equilibrium model of the Dutch economy with the capability of welfare analysis of
policy reforms. It is maintained by the CPB Netherlands Bureau for Economic Policy Analy-
sis, an independent research centre. In addition, the Ministry of Social Affairs and Employ-
ment has a dynamic microsimulation model, which can simulate ageing in a static, as well as
a dynamic way, although in practice, the ministry primarily focuses on the static ageing ver-
sion of the model and applies it in analyses of poverty and income distribution. The ministry
also has a recent special-purpose dynamic microsimulation model, SADNAP (Social Affairs
Department of the Netherlands Ageing and Pensions Model), also an in-house development.
In Finland, like in Belgium, it is a specialised institution, the Finnish Centre for Pensions,
which is in charge with pension projections. The approach is similar to that of the Danish and
Dutch administrations: a combination of a sophisticated standard model and a smaller mi-
crosimulation model for special purpose projections. Benefit data are generated by the Centre
itself; contribution data come from Statistics Finland.

Microsimulation models as primary modelling tools

France may be best classified somewhere between the complementary and the primary role of
microsimulation for pension projections. These exercises have been coordinated by the Pen-
sions Advisory Committee (Conseil d'Orientation des Reraites, COR) since 2000. National
projections are based on projections made separately by all the major pension schemes with
their own tools, on the basis of some common assumptions provided by the coordinating
agency. The COR also uses a standard model managed by the Ministry of Social Affairs for
some macro variants (for instance variants concerning demographic trends, unemployment
rates). Microsimulation currently plays a role in these projections at two levels. The PRISME
model produces the projections for the most important of all these schemes, the CNAV. In
addition, the COR has used the DESTINIE model for building its global assumptions con-
cerning the potential impact of the 2003 reform on retirement ages, the reason being that
DESTINIE was the only model with a behavioural module. Both DESTINIE and PRISME are
also used by the COR for exercises outside core projections. For instance, DESTINIE has
been used for simulating consequences of changing rules for survivors benefits, for benefits
linked to number of children or for variants concerning bonuses for people who postpone re-
tirement.
In Sweden and in the United Kingdom, microsimulation is given an even more important role.
These administrations show some special characteristics. First of all, the microsimulation pro-
cedure is highly institutionalised. The responsible ministries assign significant resources and
separate a special unit for this job. These units handle various simulation instruments. Pension
modelling can be part of general purpose models of life-cycle financing or of tax-transfer
simulation. The instruments process large administrative datasets and/or household surveys,

31
which are duly matched. They are applied in budget forecasts as well, not only in distribution
analyses.
Sweden has a long tradition of using microsimulation in pension projections. The predecessor
of the currently used MiMESIS model was put to work in 1973. This instrument is specialised
to pension analysis, in particular budget projections. SESIM, however, was developed as a
model of student loans and it was only later extended to become a tool in pension analysis, or,
for that matter, in various analyses with regard to life-cycle financing. The model is cross sec-
tional with the explicit capability of matching household members. It covers a whole range of
processes with a large number of stata among which transitions are simulated, such as educa-
tion (upper secondary: drop-out, completing; university: drop-out, completing; labour market:
to university or adult education; adult education: to university); marital status (cohabitation,
marriage, separation, divorce); change in number of children (birth, adoption, child leaving
home); labour market (unemployment, employment, miscellaneous stata), retirement (mod-
elled endogenously based on optimisation of pension wealth), wealth accumulation, housing.
The latest version also includes health. This long list reveals that SESIM is capable of simu-
lating a large variety of potential life-paths.
PENSIM2 of the UK administration was developed by the Department of Work and Pensions
(DWP) between 2000 and 2004 and has been maintained and further developed since then.
The DWP Model Development Unit has 19 full-time equivalent staff working on the Depart-
ment’s major microsimulation models, of which 3.5 undertake the maintenance and ongoing
development of PENSIM2. MDU is a multi-disciplinary unit and has contained economists,
statisticians, operational researchers, systems analysts, and social researchers. MDU models
also include a static microsimulation model of the tax-benefit system, a behavioural model for
estimating labour supply responses to various tax and benefit policies, and a dynamic mi-
crosimulation for forecasting disability benefits caseload and expenditure. MDU cooperates
on a regular basis with academics working in this field and with other microsimulation model-
lers.

32
4. An inventory of pension projection models

In this section we give an inventory of the 44 pension projection models applied by 26 gov-
ernments of the European Union. The descriptions below largely benefited from comments by
the national delegates to the ISG for which the PENMICRO team is sincerely grateful.

Belgium
MALTESE
MALTESE (Model for Analysis of Long Term Evolution of Social Expenditure) is a socio-
demographic and budgetary model, covering the receipts and benefits of different social secu-
rity schemes. Besides pensions (for wage-earners, self-employed and civil servants, public
enterprises schemes and assistance schemes), the model covers a wide range of other health
care (acute care, long-term care), social (e.g. disability allowances), family and education ex-
penditure items. Indicators produced by the model range from aggregate fiscal indicators to
average benefits by social status and type of career. MALTESE has been developed and main-
tained by the FPB (Federal Planning Bureau).
MALTESE is a system of interconnected models, comprising sub-models such as PENSION
(wage earners’ pensions), MoSES (self-employed people’s pensions) and PUBLIC (civil ser-
vants’ pensions). These sub-models are deterministic, cohort-based standard simulation in-
struments based on administrative datasets. Generally, the old-age retirement rate at 65 for
males is assumed to be constant. The female ‘coverage’ rate is held constant at 65 (coverage
equals own pension entitlement, survival pension and the cases where the husband has a pen-
sion for both spouses). The entry probabilities into the disability scheme have been calculated
from the exogenous labour force projection (number of potentially active people). The aver-
age amount of newly-granted old-age pension is estimated for a number of typical careers
based on specific career profiles. The general average pension amount per scheme (employ-
ees, self-employed, civil servants) is weighted by its share in the total number of pensioners of
each type. The income distribution is supposed to remain constant in the projection.

MEP
MEP (Microeconomic Pension Model) is a typical agent standard simulation model based on
the option value approach. The model has been prepared by the FPB (Federal Planning Bu-
reau) in 2006, using the SAS macro language. Four hypothetical private sector employees
(male and female blue and white collar workers) are modelled, who are deemed to be repre-
sentative for the purpose of examining actuarial neutrality of Belgian first pillar pensions (in-
dividual replacement rates, changes in pension wealth, implicit tax, option value). MEP calcu-
lates for every year in which the individual is eligible for the pension benefit the flow of ex-
pected future benefits until death. All benefits are then discounted back using a discount rate
and a survival probability. As regards data on past individual income, the model includes a
matrix containing wages per day for every combination of wages and labour market status, for
all ages between 20 and 64 and for all the years between 1955 and 2004. To keep things sim-
ple, individuals are assumed to be either singles or the sole income earners in the household.
It is assumed that the four typical employees were born in 1940 and entered the labour market
at the age of 20.

33
MIDAS_BE
MIDAS (an acronym for ‘Microsimulation for the Development of Adequacy and Sustainabil-
ity’) was developed as part of the Adequacy of Old-Age Income Maintenance in the EU (AIM)
6th Framework project in order to simulate the adequacy of pensions in Italy, Germany and
Belgium. MIDAS is a dynamic microsimulation with dynamic cross-sectional ageing, i.e. it
starts from a cross-sectional dataset representing a population of all ages at a certain point in
time. Estimates of behavioural equations in the Belgian version of the model are based on
individual survey data of the Panel Set of Belgian Households. Other survey data are also
incorporated into the instrument, e.g. educational attainment is determined on the basis of
observed probabilities from the Labour Force Survey. In Germany, the model is based on the
German Socio-Economic Panel (GSOEP); in Italy it is based on the European Community
Household Panel (ECHP) linked to other datasets.
MIDAS was produced by three public research institutions in close cooperation (FPB, DIW
and the Italian ISAE), with FPB as project manager. It has a modular form consisting of a
demographic module (with four sub modules, namely the birth process, the survival process,
the education process and the marriage market, a labour market module and a country-specific
pension module. The demographic and labour market modules are common between all three
countries, though behavioural equations are naturally estimated and calibrated on country-
specific data.
Individuals are born, go through education, marry or cohabit, enter the labour market, divorce,
retire and, finally, die. All these main events in a lifecycle are simulated by the model. Transi-
tion probabilities for a number of these processes are produced by logit regressions (e.g. mar-
riage market, labour market position). Assumptions of the model are aligned to those of the
Ageing Working Group, wherever possible, with the help of econometric alignment proce-
dures. The model therefore aligns to AWG demographic and macroeconomic projections, and
generates a wide variety of indicators of income inequality, poverty and (re)distribution, both
for workers and retirees, and potentially also indicators of nonneutrality.
The model is written in the programming language LIAM, which was designed by Cathal
O’Donoghue (TEAGASC) and further developed by the Belgian Federal Planning Bureau
(FPB) and the German Institute for Economic Research (DIW).
MIDAS is in the final development phase and it will deliver its first results in November
2008.

MIMOSIS
MIMOSIS (Microsimulation Model for Belgian Social Insurance Systems) is a static mi-
crosimulation model without behavioural responses for social security and personal income
taxes in Belgium, programmed in FORTRAN 95. The model has been developed by the Uni-
versity of Antwerp, the University of Leuven the University of Liège and the FPS Social Se-
curity in close collaboration with the Crossroads Bank for Social Security who provided the
input data.
MIMOSIS runs on a sample of 305.019 individuals, which is about 3 percent of the total
population. The sample has been constructed by sampling at random 100.000 individuals
from the National Register. After this, all other household members of the sampled individu-
als have been included in an extended sample. Weights have been constructed to aggregate
results at sample level to the level of the overall population. For all individuals in the ex-

34
tended sample, administrative data have been taken from the Datawarehouse Labour Market
and Social Protection, which is maintained by the Crossroads Bank for Social Security.
This sample with administrative data is exploited to program the legislation, applicable to the
most important social security policy domains, in different modules. The final tool then al-
lows for the (re)calculation of benefits received and taxes paid for each individual and/or
households in the dataset. The modules cover the following areas: contributions, family al-
lowances, sickness and disability benefits, unemployment benefits, pensions, personal income
taxes, social assistance benefits plus a module for the assessment of budgetary and distribu-
tional impact of reforms. In order to assure great flexibility, a large number of external vari-
ables have been endogenised, i.e. calculation rules of the tax-benefit system have been param-
eterized to allow for the computation of disposable income.
Simulation possibilities for pensions are currently restricted to changes in observed pensions
only (so called ’welfare adaptations’, one-off benefit increases). Changes in the benefit for-
mula cannot be modelled, as of yet but a more elaborate module for calculating pensions
based on past career information is currently being developed and shall be integrated in a next
version of MIMOSIS.
The evaluation tool covers and models the majority of policy relevant income concepts, ex-
cept for income from property such as imputed rents and income from capital. Mortgage in-
terest payments and housing allowances are also excluded from modelling.
Currently MIMOSIS offers the possibility to change the policy parameters of the 2001 legisla-
tion but parameter files of more recent tax-benefit years can easily be integrated. Changes in
the legislation beyond changes in parameter values, e.g. a whole new way of calculating so-
cial security benefits, cannot be simulated without changing the source code. In the pension
module, the effects of benefit increases can be investigated according to type of pension, type
of scheme (wage-earner, self-employed) and date of pension award.
The main outcome of the model is a comparison of net disposable incomes before and after a
reform. When evaluating distributional consequences of a reform, the user can choose from
among a number of poverty benchmarks, equivalence scales, units of analysis. The basic unit
of analysis is the individual but also households can be used as unit of analysis. Apart from
tables per decile of equivalised income of losers and gainers, poverty incidence and intensity,
replacement rates and various measures of income inequality, results are also produced for
different socio-economic classifications and age groups. The model also calculates at-risk-of-
poverty rate by age and gender, by most frequent activity and by household type.

Bulgaria
ILO PENS (BG)
The Bulgarian government applies an upgraded version of the ILO PENS, a sub-program of
the ILO Social Budget Model, tailored to the needs of pension budgeting in Bulgaria. It is part
of the annual process of preparing the budget for the general government but it is also in use
for medium term projections and for analysing short term budgetary consequences of policy
changes.
The ILO PENS (BG) covers old-age pensions, disability benefits and survivors’ benefits. It is
managed and maintained by the National Social Security Institute (NSSI). In terms of organi-
sation, it is part of the general social budgeting process. The authorities employ three people
working on the social budget model and another three working on the pension sub-program.

35
The ILO PENS is a sub-program of the ILO Social Budget Model developed and maintained
by the ILO FACTS, the International Financial and Actuarial Service of the ILO. It is a stan-
dard deterministic cohort-based projection model. For each generation, the transition of an
insured person (active, inactive and pensioner) is mapped into the next year’s status by using
actuarially assumed transition probabilities (mortality rates, incapacity rates, retirement rates,
etc) and applying eligibility conditions and pension formula. The ILO PENS is written in Vis-
ual Basic with an Excel face. Its aim is to give short-term projections on key aggregate vari-
ables such as the number of beneficiaries and the overall expenditures, revenues and deficit of
the system.

Czech Republic
Czech Pension Model
The Czech Republic has standard simulation pension models, which enable both typical agent
and cohort-based modelling. The instruments have been conceived by an expert team in
2004/05 and since then maintained and updated by the Ministry of Labour and Social Affairs.
The models were applied in the evaluation of various parametric and paradigmatic reform
proposals. For the AWG projections a model based on similar principles and methodology is
used.
Both modules (typical agent and cohort based) are built on administrative datasets of the so-
cial security authority, which are complemented by survey data on labour force characteristics
as well as census data on family features.
Beyond age, gender and employment type (employed/self-employed), old-age benefit projec-
tions are also disaggregated on the basis of income ranges (10 per gender). When assessing
newly granted pensions, cohort members are classified into several groups based on their av-
erage earnings and career lengths. However, the model is currently not able to realign these
parameters to prospective changes in the projected labour market situation.
The main outcomes of the cohort-based module include various fiscal indicators (expendi-
tures, revenues, accumulated surplus/debt), an overall replacement ratio and the share of new
retirees under the poverty line. The typical agent module calculates individual replacement
rates, internal rate of return and implicit tax rate.
Regarding datasets, an administrative database on pension accruals is currently in the process
of preparation, as part of setting up individual accounts for administrative purposes.

Denmark
Danish Pension Model
The Danish Pension Model used in the context of the Ageing Working Group was developed
and is maintained in-house. It is an actuarial valuation model taking the standard cohort-based
approach, projecting public pensions. The projection does not only takes into account age,
gender and labour market position in the exercise, but also ethnic origin (native, immigrant,
descendant of immigrant).
Assumptions on the future age and sex composition and the corresponding mortality and fer-
tility rates are taken from Eurostat. The ethnic/racial composition is assumed constant. Inter-
nal consistency of macro variables such as wage growth, the real rate of return, employment
and inflation is ensured by the production function approach of the AWG, which sets the la-
bour share of GDP constant. The basis for the projection is the RAS administrative survey.

36
The model is not suited to produce micro level output, as results are limited to the number of
contributors/beneficiaries to the schemes, and aggregate revenues and expenditures.
The complement of the model is a microsimulation tool, called LAW (see below in detail).
This reflects the modelling strategy of the administration, which is based on a cohort model
for measures on long-term sustainability and an “add-on” model, which can answer questions
on the distribution of economic resources.

LAW
Denmark’s LAW model was developed and is operated by the Danish Ministry of Finance. It
is currently in use, while it is also being revised and continuously updated. It is part of a larger
model, which also include, among others, a tax model and a model for housing subsidies.
LAW is a static microsimulation model, which run in SAS with a background of a complex
panel dataset from Statistics Denmark, by the matching of 25-30 different administrative mi-
cro databases. The separate datasets are matched by the identification number of Danish resi-
dents. Although these data are individual they allow the reconstruction of households as
household members (sometimes by assumption) can be matched.
Formal calibration for aggregation is from the Danish Pension Model (DPM; see above). The
administration considers LAW a complementary model to the DPM, which can answer ques-
tions on the distribution of the economic resources.
Calculations are carried out on a random sample. The dataset contains retrospective informa-
tion on employment and earnings back to 1979 at the maximum. Accruals in the funded
scheme are also available: around 90 percent of funded schemes are covered by the sample.
The remaining 10 percent has been 'individualised' by using a simplified matching technique.
Other types of available micro data include information on health conditions, social condi-
tions, use of day-care centres, economic relationship to public sector, demographic conditions
and education. All data inputs are on individuals, while households are treated separately in a
typical agent fashion.
One notable feature of the model is that in the determination of labour supply, "dynamic" or
second-order effects of tax changes with assumptions on gender-specific labour supply func-
tions are used to derive behavioural responses.
The model produces a wide range of output indicators including first pension / last wage type
of replacement rates, which are more telling than the standard average pension / average wage
ratio. It is also capable of producing the distribution of pension wealth and future pension
annuities payments.
The model is now being developed for long-term projections but it will still remain a static
microsimulation model.

Germany
AVID
AVID (Altersvorsorge in Deutschland, Retirement Pensions Provision in Germany) is a re-
peated study based on large sample surveys of clients of the German social pension insurance.
The survey was first undertaken in 1996 and repeated in 2002. The second round was ex-
tended in 2004. It was commissioned by the Federal Ministry of Labour and Social Affairs

37
and the Federation of German Statutory Pension Insurance Institutes and it was conducted by
an external contractor.
The 1996 round covered a sample of over 12,000 people of the German resident population
between the ages of 40 and 60 (birth cohorts from 1936 to 1955) having a social security re-
cord and eligible for old-age pension at the age of 65. It was extended to the spouses of re-
spondents as well, irrespective of the age and nationality of the former. The questionnaire
design covered a whole range of potential income sources during old age, such as statutory
pension insurance, private and public supplementary systems, civil servants‘ pension scheme,
farmers‘ old-age pension scheme and schemes for the independent professions. It also in-
cluded private provisions, such as life insurance and private pension insurance, property. The
2002/2004 round has a sample size of close to 14,000 people and their spouses.
The surveys collected rich sets of information on the family and household background, eligi-
bilities to alternative future income sources upon retirement as well retrospective data affect-
ing entitlements. This information was matched with the personal pension accounts held by
the social security system including a detailed retrospective labour market and wage career.
This combined set was extrapolated by a dynamic microsimulation model. The model focuses
on transitions across labour market stata but puts aside other socio-economic considerations
such as e.g. changes in the marital status. The list of labour market stata is highly detailed
including 14 different stata ((1) employed, subject to social insurance contributions, 2) child
care, 3) housekeeping, 4) nursing relatives in need of care, 5) long-term sickness, 6) unem-
ployment, 7) old-age pensioner, 8) unpaid work in family businesses, 9) low-paid employ-
ment, not subject to social insurance contributions, 10) civil servant, 11) self-employed, 12)
limited earning capacity, 13) education, 14) other). Working hours and income position are
also modelled for the relevant stata. Transitions between labour market states and working
hours are modelled by hazard rate models; self-employment and working hours entries are
modelled by logit/probit; the wage model is based on linear regression.
The reference period of extrapolation started by 1992, that is, it covered the years 1992-1996
in the first round and 1992-2002 in the second round.
The application of the model is the extension of the scope of standard models, that is, to go
beyond aggregate cash flows and also identify risk groups and simulate the impact of various
reform measures on the lower edge of the income distribution. It also serves in pension policy
evaluation and planning and calculations on the link between work histories and old-age in-
come. In addition, AVID also has the potential to enlarge the picture from the individual to
the household and take into account alternative sources of income after retirement other than
old-age pensions, such as other retirement provisions and withdrawal of private savings.

German Pension Model

The regular procedure of projecting future pension expenditures is managed by the Ministry
of Health and Social Security in collaboration with the German Pension Insurance. The pro-
jection proper is made by an expert team including representatives of these two institutions
and people from the German Federal (Social) Insurance Authority but the consultation proc-
ess also involves experts of the Federal Ministry of Economics and Labour and the Finance
Ministry.
The German Pension Model is a standard semi-aggregate model, which works with averages
for cohorts and sexes separately for each of the three main provisions, old-age, disability and

38
survivors. It quantifies the demographic impact on future pension expenditures and applies a
partial equilibrium model in order to calculate the effects on contributions and pensions.

Estonia
Estonian Long-Term Pension Budget Model
The Estonian long-term pension budget model was developed and is operated by the Finance
Policy Department of the Ministry of Finance. The model consists of Excel spreadsheets, with
a partial equilibrium, "top down" structure. It is a deterministic macro simulation model based
on cohorts, encompassing all pension schemes except the voluntary pillar. The unit of projec-
tion is the cohort, which means that the population is differentiated by age and gender. An
important feature is that this is done not only in terms of labour market position (and retire-
ment status) but an age and gender-specific relative wage structure is also incorporated into
the model.
The model contains one important simplification, in that new pensioners are not projected
separately from the existing pensioner stock in the disability scheme, but an evolution of the
stock is assumed, calculated using initial (stock) data and an (arbitrary) change vector. In the
old-age schemes, retirement is modelled by the adjustment of retirement age profile based on
changes in participation rates of older cohorts, so retirement behaviour is exogenous and
manually driven.
The assumptions on the future age and sex composition of the population (including mortality
and fertility rates) are taken from Eurostat and are all exogenous, while the ethnic/racial com-
position is not explicitly included in the model. Macro variables (wage growth, real rate of
return, employment, inflation) are all exogenous, coming from the commonly agreed Eurostat
set of assumptions, and, like in other cases of the AWG projections, the wage share remains
constant during the projection period so as to guarantee internal consistency of these vari-
ables.
The Estonian long-term pension budget model is a semi-aggregate model applied in projec-
tions of future revenues and expenditures.

Ireland
Irish Pension Model
The Irish pension model was developed and is operated by the Ministry of Finance and is
regularly updated. The model is described as a partial equilibrium standard simulation deter-
ministic model based on cohorts, encompassing public service pensions, social security pen-
sions and assets of funds and reserves (the National Pensions Reserve Fund). New pensioners
and their benefits are not projected separately from the existing pensioner stock, but an evolu-
tion of the stock is assumed. This differentiation between “stock” and “entry” pensioners and
pension levels leaves some scope for possible development in the Irish model.
Assumptions regarding demography and macro variables are in line with the Eurostat and
AWG projections.
The model is not built for producing micro level output but aggregate revenues and expendi-
ture and the number of contributors / beneficiaries to the schemes. That said, an important
feature of the model is that effects of pension reforms on the social welfare budget can be
incorporated into the model.

39
Greece
Greek Pension Model
The National Actuarial Authority performs calculations on theoretical replacement rates in
line with ISG methodology. Simulations incorporate cases where people have just entered
employment or social insurance or will join the labor force in the near future or became in-
sured for the first time only a few years earlier. The fund modeled (IKA-ETAM) covers all
private sector employees (about 55 percent of the employed), therefore the modeled life-paths
can be deemed as typical. Modelling work related to the AWG exercise is currently under-
way.

Spain
Spanish Pension Model (Modelo de proyección de pensiones del IEF)
The Spanish pension model was developed and is operated by the Ministry of Finance. This is
a deterministic cohort-based projection model, covering old-age, early retirement, survivor
and disability pensions for private employees and self-employed as well as minimum pension.
A separate module covers public sector pensions, including disability, early, minimum and
war pensions. The model is operated in Matlab and is updated on a yearly basis.
The unit of projection is the cohort, yet the simulated characteristics are not limited to age and
gender, but the model also differentiates within cohorts by labour market position (making or
not making contributions) and pension scheme but not by level of education. New pensioners
and their benefits are projected separately from the existing pensioner stock, which helps to
explicitly capture retirement and the dynamics of benefits over time. New disability pension-
ers are projected based on past experience. Minimum pensions are assumed to develop in line
with the average pension.

France
DESTINIE
The French DESTINIE model was developed and is operated by the French Statistical Office,
INSEE. It is updated on a regular basis: whenever changes to the legislation occur, or when a
new version of the underlying dataset is available, the DESTINIE model is revised. A more
complete rewriting is currently under way, under the name DESTINIE II. The new model
essentially differs from the previous one in terms of programming language (Perl replacing
Turbo Pascal), modularity, options of retirement behaviour, and consistency with macro-
projections. But the main choices and functionalities of the first version have been kept. Even
if Destinie II is not fully operational at this stage, it progressively replaces Destinie I and the
details below refer to this new version.
DESTINIE is a dynamic microsimulation model based on a household survey, the Financial
Assets Survey (Enquete Patrimoine, 2003) covering a 65,000-person representative sample of
the population. The model covers the functioning of France’s entire pension system including
widow pensions, family supplements, minimum pensions and a means tested minimum. But
small specific schemes are not simulated in details. They are aggregated to the major schemes
they are the closest to. For instance, wage earners in large public firms that often have specific
schemes are aggregated to the group of State employees and are assumed to share the same
pension rules.

40
Individuals in the model are characterised by their marital status, age and gender. Their ca-
reer paths in terms of labour market status until 2003 are also included in the dataset. Wages
are simulated through the model, thus also when creating the retrospective career histories,
based on wage equations estimated on fiscal data matched with the LFS (the so-called Fiscal
Income Survey). After the base year, the model simulates both demographic and economic
trajectories from the initial year up to 2050, using a mix of deterministic rules, behavioural
hypotheses and random drawings. Demographic developments are exogenous, but highly so-
phisticated, including the simulation of the school-leaving age, first "marriage", recoupling,
rupture, and children born from all these unions. The model applies differentiated mortality
rates depending on age on leaving school. It provides detailed information on household
structure, including size and combined income. It also simulates kinship ties between non-
core people, for instance the survival of parents or siblings. This is done through a closed
population approach. Pseudo-kinship ties are reconstituted within the sample at the outset and
then updated or generated by assuming that all unions take place within the sample, with a
renewal of the sample ensured by births resulting from these unions.
As for the development of earnings and pension accruals, equations (transition probabilities)
are estimated by gender and age on leaving school using data from the Labour Force Survey
and other sources.
The predicted probabilities for demographic or labour market events are adjusted in order to
fit every year the long-term demographic projections and the official labour force forecasts
(labour force participation rates by gender and five-year ages) made by INSEE and to be con-
sistent with the assumptions made within the macroeconomic environment. Conditional on
labour market status, annual wages are finally imputed using the same wage equations as the
ones used for the retrospective reconstitution of careers.
In projecting retirement decisions, the model incorporates five options to determine the choice
of retirement age. The first one assumes that individuals retire as soon as they reach the so-
called full rate (i.e. what is generally considered as the normal pension), access to the full rate
depending both on age and the number of years of contribution through a rather complex non-
linear formula. The two next options assume some form of utility maximization with an in-
come-leisure trade-off, either instantaneous (people retire as soon as the value of leisure out-
weighs the immediate income loss) or prospective (the option value model of Stock and
Wise). The two last options are based on measures of social-security wealth (SSW), i.e. the
actualised expected flows of benefits from retirement age until death, with two sub-options: a
myopic one under which people postpone retirement if a one year postponement increases
their SSW and a more forward-looking one analogous to the Stock and Wise model where
people delay if delaying leaves them with the option of getting a higher SSW level at a later
age.
The model is able to produce pension distributions both in terms of retirement age and retire-
ment income. It is also possible to extract poverty measures from it, such as the number of
elderly falling below the 60 percent poverty line.

PRISME
The PRISME model of the French pension system is utilised by the Conseil d'Orientation des
Retraites (Pensions Council) within the EU’s Ageing Working Group projections, among oth-
ers. The model is a dynamic microsimulation model based on the administrative data from the
General Scheme for Employees CNAV. Its time horizon is 2050, with the base year 2006 and

41
is updated every 2 years. The tool covers the whole population of the entire pension system1
registered for Social Security, contributing or not to the General Scheme.
The model reconstructs entire life paths, including school completion, household structure and
kinship ties, the completion of the child-raising period and differentiates between seven la-
bour market states 1) labour market participation CNAVTS, 2) labour market participation
CNAVTS-analogous, 3) labour market participation non-CNAVTS, 4) unemployment 5) sick
leave, 6) disability, 7) other. Thereafter, individuals end their working lives and die.
The main objective of the model is to forecast for the medium or long term the resources and
the expenses of the General Scheme. In addition, PRISME allows forecasting at an individual
level the accumulated entitlements to the Old Age Insurance Scheme, professional paths,
wages of the private sector workers and the amount of pension (direct entitlement and widow
pensions). Main published results concern general outcomes for the COR report 2006 and the
update of 2007, detailing: the number of persons contributing (2003-2050), the number of
pensioners (direct rights and widow pensions), the total amount of contributions and the bal-
ance of the General Scheme.
As for more micro level output, the model is able to produce the number of retirees at the av-
erage age of retirement, the number of early retirements (anticipated pensions) in the General
Scheme, the share of pensioners concerned by the reduction in the case of early exit (decote)
and by the supplement in case of deferment (surcote), the average pension benefit at retire-
ment age by sex.

Italy
CAPP_DYN
CAPP_DYN is a dynamic, cross-sectional, population based, closed microsimulation model of
the Italian population, developed by Centro di Analisi delle Politiche Publiche (CAPP), a re-
search centre for the analysis of public policy, run jointly by the Universities of Modena and
Bologna. CAPP_DYN is used for modelling various policy options in the field of social secu-
rity (not just pensions).
The primary database of the model is the Bank of Italy Survey of Household Income and
Wealth (SHIW_02), comprising 8001 households and 21,400 individuals. For the purpose of
statistical adjustments, census data were used. Additional survey data have been taken from
various sources to simulate educational choices, earning equations and labour market transi-
tions. Special routines calibrate the model’s endogenous results in a way to bring them in line
with the exogenous basic economic hypotheses (per capita income growth, real earnings
growth).

1
– Régimes des salariés: CNAV, régime des salariés agricoles, ARRCO, AGIRC, IRCANTEC. – Régimes des
indépendants: ANCAVA (artisans), ORGANIC (commerçants), régime des exploitants agricoles, CNAVPL
(professions libérales), régime élémentaire obligatoire de la CANCAVA, régime complémentaire obligatoire de
l’ORGANIC, régime complémentaire obligatoire des exploitants agricoles, régimes complémentaires obliga-
toires de professions libérales (CARCD, CARMF, CARPIMKO, CAVP). – Régimes de fonctionnaires : régime
de la fonction publique de l’État, CNRACL (collectivités locales), régime additionnel de la fonction publique. –
Régimes spéciaux: Banque de France, CNIEG (industrie électrique et gazière), CRPCEN (clercs et employés de
notaire), ENIM (marins), FSPOEIE (ouvriers de l’État), régime de retraite des mines, SNCF, RATP.
http://lesrapports.ladocumentationfrancaise.fr/BRP/064000302/0000.pdf

42
The unit of simulation is the individual, but data on family structure and its changes are re-
corded, too. All individuals in the sample are involved in a considerable number of demo-
graphic and economic events, such as birth, education, marriage, work, retirement and death.
Economic and demographic transitions among states are simulated by Monte Carlo processes.
A set of matrices and econometric models are employed for generating transition probabili-
ties. The CAPP_DYN model has a recursive structure consisting of a set of modules executed
in a predetermined order (demographics, educational choices, job decisions and earnings es-
timation, retirement decisions).
The model covers most social security benefits of the Italian welfare system (old-age, sur-
vival, disability, minimum pension) and it is capable of producing a large number of indica-
tors on intra-generational as well as intergenerational distributive effects (replacement rate,
internal rate of returns etc). A comparison between labour and social security income is also
possible. Since the model does not cover all income components (real and financial capital
income are still missing), a complete analysis of income distribution and poverty in the long
run is not possible. Recently a new module has been added in order to estimate the likely evo-
lution of the frail part of the population in the medium to long term and the cost of the intro-
duction of a long-term care public programme. At the same time, the research group is cur-
rently working on the substitution of the primary database, the SHIW, with the Italian EU-
SILC 2005 survey, which would allow a better comparison with other European models.

CeRP models
The Centre for Research on Pensions and Welfare (CeRP) in Italy developed a family of pen-
sion models including a typical agent model a semi-aggregate model and a micro-simulation
model. This short summary is based on the institute’s annual report.
The typical agent model computes pension benefits and money’s worth measures for a set of
representative Italian workers according to the rules of the legislation in effect. By aggregat-
ing individual results, it evaluates the impact of alternative pension rules on pension expendi-
ture.
The semi-aggregate model (CeRPSAM) was elaborated in cooperation with the Dutch Social
and Cultural Planning Bureau The inputs of the model are country-specific demographic and
household projections (mainly from Eurostat data), assumptions on future labour participation
rates and labour productivity growth, and the current income/benefit positions of the popula-
tion. The model produces estimates of future developments on a number of economic key
variables, and of indicators for the future income/benefit positions of specific groups. The
latter are subsequently implemented in European Community Household Panel Survey
(ECHP)’s micro data through a weighting procedure. This makes it possible to calculate indi-
cators of inequality, redistribution and poverty for the year 2025.
The microsimulation model (CeRPSIM) is able to simulate heterogeneous lives and earnings
histories for individuals belonging to various cohorts, and to compute the resulting pension
benefits. The simulated population evolves throughout a set of deterministic and stochastic
elements. Discrete-state changes (marital status, labour status, etc.) are conditional on indi-
vidual socioeconomic characteristics and are modelled throughout a Monte Carlo procedure.
The process for the lifetime earnings paths is modelled, for each individual, as the sum of a
group-specific deterministic component and of a group-specific stochastic component, esti-
mated from a sample of administrative data.

43
RGS pension model
The pension model of the Ministry of Economic and Finance Department of General Ac-
counts (Ragioneria Generale dello Stato) maintains a cohort-based standard simulation model
based on administrative data. The model is used to assess the aggregate financial effects of
reform proposals and serves as the underlying instrument of the AWG projections of public
pensions. Sub-groups in the model are distinguished by age, sex, benefit type, source fund
(private/public), contributor type (active/dormant/pensioner) and type of regime (earnings
related, contribution based, mixed). Technically, changes status are handled by a transition
matrix, in which probabilities have been derived from past data. Up to age 42, the number of
new contributors is set equal to the increase of employed people within the cohort. The model
is quite flexible to take on board exogenous macroeconomic projections except for exit from
the labour market where probabilities have been estimated endogenously from past trends and
eligibility criteria and only approximated to aggregate external forecasts. Projections on vari-
ous sub-groups are aligned to macro level projections (e.g. in the case of wages) by using
multipliers. Non-linearities of the pensions system (e.g. minimum pension or different rates of
benefit indexation based on benefit amount) are also dealt with through an index of variability
(variation coefficient) and a distribution function. The model makes use of historical data
about workers with a contribution past, including dormant members who are no longer con-
tributing but will be able to apply for a pension later.

Cyprus
ILO PENS (CY)
The ILO PENS is a sub-program of the ILO Social Budget Model developed and maintained
by the ILO FACTS, the International Financial and Actuarial Service of the ILO. It is a stan-
dard deterministic cohort-based projection model. For each generation, the transition of an
insured person (active, inactive and pensioner) is mapped into the next year’s status by using
actuarially assumed transition probabilities (mortality rates, incapacity rates, retirement rates,
etc) and applying eligibility conditions and pension formula. The ILO PENS is written in Vis-
ual Basic with an Excel face. Its aim is to give short-term projections on key aggregate vari-
ables such as the number of beneficiaries and the overall expenditures, revenues and deficit of
the system.
Cyprus uses the ILO PENS. The model has been customised in order to closely comply with
local legislation. For further details on the model see the corresponding paragraph above.
The ILO PENS (CY) is operated by the Social Insurance Services of the Ministry of Labour
and Social Insurance. It is updated in every three years, in order primarily to incorporate
amendments to legislation and results from experience analyses. Its results are used in actuar-
ial valuations, long-term budgetary planning, modelling pension reform options and in the
cash-flow projections between the Consolidated Fund and the Social Insurance Fund.

Latvia
Latvian Pension Model
The Latvian pension model (LPM) is run under the auspices of the Ministry of Welfare. It is
written in Visual Basic and has an Excel user face. It is applied to a variety of public cash
programs such as the old-age pension, disability benefits, survivors’ benefits as well as other

44
functions such as short-term sickness, work injury, unemployment, maternity and funeral
benefits.
The LPM is a standard semi-aggregate model. The level of disaggregation is the cohort (year-
group) and sex. Projections are based on participation and wage profiles by age and sex pro-
vided by the State Social Security Agency (SSSA) and on data on current pensioners and pen-
sion profiles from CSB and SSSA. Assumptions regarding developments of major macroeco-
nomic trends are in line with the assumption set of the AWG. Reweighing of the cohorts, that
is, the imitation of future developments of the retirement process, more specifically, the aver-
age number of new retirees and the pensioner career of cohorts, is based on assumptions.
Although the LPM is a cohort model, it is also used to draw typical age-earnings careers in
order to calculate theoretical replacement rates. Nevertheless, the main application of the
model seems to be the projection of financial developments, aggregate spending on old age,
disability, short-term sickness, work injury, unemployment, maternity, survivors, funeral
benefit and even administration costs. A special section of the LPM is modelling the conse-
quences of the switch of the system from a defined benefit pension formula to a non-
financially defined contribution (NDC) system.

Lithuania
PRISM (LT)
PRISM (Pension Reform Illustration and Simulation Model) is a general-purpose pension
model developed by Patrick Wiese. In Lithuania it is maintained by the Ministry of Social
Security and Labour and it contains state social insurance pensions covering old-age pensions
(including mandatory private funds), disability benefits and survivors’ pensions.
PRISM is a semi-aggregate model. It can operate as an “average person parameter” model or
as a distributional model. In this case new retirees are assigned a bivariate normal distribution
of lifetime wages and of pension service. The sub-groups are created by age and sex. Demo-
graphic projections are made by the model or, alternatively, an exogenous population projec-
tion can be imported into the model; the assumptions are in line with those of the AWG. The
model has the capacity to endogenise the calculation of GDP, that is to say, to derive it from
inflation, wage growth and employment growth, thereby ensuring internal consistency.
The model calculates the usual aggregate measures, such as the cash-flow of the system, asset
accumulation in the pre-funded pillar, and the number of contributors and beneficiaries. In
addition, it produces entry replacement rates for men and women with 35 years of service
assuming various wage levels (50 percent, 100 percent and 200 percent). In addition to these,
PRISM is able to generate measures for lifetime internal rates of return for the various sub-
groups as well as cumulative measures such as the implicit debt of the system.

Luxemburg
LuxMod
LuxMod is a Computable General Equilibrium Model (CGE) for the Luxembourg economy
and social protection, managed by the Inspection générale de la sécurité sociale (IGSS). The
instrument has been programmed in GAMS and updated on an annual basis. The model dis-
aggregates the economy into four sectors: households, enterprises, government and rest of the
world. The instrument covers all social security schemes but does not simulate the acquisition
of pension rights. The model is based on aggregate administrative data. It has been used in the

45
annual budgeting process and during the preparation of social security reforms and from
2008, LuxMod will be used for updating the Luxembourg Stability and Convergence Pro-
gramme. According to plans, it might be linked to a microsimulation model in the future.

REDIS project
The REDIS project started in 2007 and aims at assessing the coherence of social transfer poli-
cies with the help of poverty and redistribution indicators, focusing on individuals and house-
holds. The project is carried out by CEPS/INSTEAD Luxembourg (Centre d'Etudes de Popu-
lations, de Pauvreté et de Politiques Socio-Economiques) in partnership with IGSS (Inspec-
tion Générale de la Sécurité Sociale) and financed by the National Research Fund. As a start-
ing point, the redistributional effects of the Luxembourg tax reforms of 2001/2002 have been
analysed by using the static microsimulation model EUROMOD. The analysis of the tax re-
form was based, on one hand, on administrative data drawn from the Luxemburg Social Secu-
rity Data Warehouse and the EU-SILC survey. The model was run with these two different
types of datasets and outputs (inequality indicators and poverty rates) compared. The unit of
analysis was in both cases the fiscal household.
The REDIS project foresees some important developments to EUROMOD. Among others, a
new module enabling explicit modelling of pension calculation rules is planned. Another
planned extension will serve the evaluation of effects of co-payments for health care services
on net income. In certain areas (such as the labour supply of women or co-payments in health
care services) EUROMOD calculations are going to be extended by behavioural equations.
The project will also propose a gradual strategy for developing dynamic microsimulation
models with enhanced capabilities in the area of social policy analysis, to be completed by
2010. Current plans include a cohort microsimulation model for the modelling of lifetime
income distribution and a pilot dynamic population model.

SOBULUX
SOBULUX (Social budget simulating software for Luxembourg) is a cohort-based standard
simulation model set up by the Inspection générale de la sécurité sociale (IGSS), based on
previous ILO-type macro-projections. The regularly (most recently in 2005) updated tool is
based on administrative datasets of the IGSS. In order to take account of peculiarities of the
Luxembourg labour market (high proportion of migrant workers), the instrument was de-
signed to include dimensions such as country of origin or employment status (beyond the
general breakdown by age, sex and benefit type). The model thus makes a difference between
total labour force and ’national’ labour force. Labour force participation rates are computed
by applying entry probabilities to inactive population or exit probabilities to active population.
The administration also applies SOBULUX to model the acquisition of pension accruals by a
separate module. The tool has been used for long-term planning, the assessment of pension
reform options and in political debates. EPC-AWG pension projections are also produced by
SOBULUX.

Hungary
Hungarian Pension Models
The National Bank of Hungary and the Ministry of Finance both have their own cohort-based
standard pension models developed and maintained internally by the respective institutions.

46
Both models rely on administrative datasets provided primarily by the National Pensions Au-
thority and carry out deterministic calculations along the dimensions of age, sex and benefit-
type. Both models (programmed in Excel VBA/Excel) draw on exogenous demographic and
labour supply assumptions but are able to adjust the latter in order to ensure consistency with
rules on pension entitlement. The instrument of the MoF is also able to take account of mor-
tality differences between disability and other pensioners. The initial development of the
model of the MoF (which will now serve as a basis for the 2008 round of AWG projections)
took approximately 2.5 man-months (including data collection and calibration). Both models
have been used to calculate various fiscal indicators in a number of policy and parameter sce-
narios. Their main weakness lies in the way exogenous (past and projected) changes in the
labour supply are reflected in the future evolution of average pension accruals. This work
could not be carried properly due to the lack of a retrospective data series on benefit accruals.
This shortcoming might be redressed following the involvement of the National Pensions Au-
thority in the new cycle of the labour market survey carried out in 2008.

NYIKA
The NYIKA model is a recent development born from the cooperation of the Pensions and
Ageing Roundtable (NYIKA by its Hungarian acronym), an independent expert body invited
by the Prime Minister, and Deloitte Touche Tohmatsu, a multinational consultancy. The
Roundtable, jointly with similar expert groups on education and taxation, is a chapter in the
budget of the Prime Minister’s Office.
The model was developed with the aim of testing the effects of various pension reform sce-
narios on financial sustainability and poverty in old age. It is a dynamic microsimulation
model with dynamic ageing. The life events modelled are entry to the labour market (with
transition from school to labour), marital stata, labour market positions (full time employed,
part time employed, unemployed, inactive below retirement age), and retirement. Models of
life event are mostly based on past and current patterns and require further refinement.
The model is based on administrative data on contributors to the pay-as-you-go pillar
(KELEN with over 6 million records), which is matched with contributory records of the
mandatory private funds of the second pillar (KPN). The model also gains further information
on current pensioners (from the NYUFUR dataset of the pension administration) and con-
tributors (from Labour Force Survey).
The first results of the model have recently been released but its current position in the ad-
ministration and its future developments are still uncertain.

Malta
PROST (MT)
PROST (Pension Reform Options Simulation Toolkit) is a general-purpose model of the
World Bank. It is a combined semi-aggregate and typical agent model, in that it derives con-
clusions regarding macro expenditure and revenue flows by reweighing cohorts and genders
based on a demographic projection, but it is also capable of tackling as many as six typical
careers defined by the user. Consequently, it overcomes the usual problem of typical agent
models of producing aggregate results and at the same time it avoids some deficiencies of
cohort models such as remaining silent about replacement rates and similar life-cycle meas-
ures.

47
Accordingly, PROST produces the average entry pension over average gross wage type of
replacement rate. Indeed, this version of the replacement rate is not without problems of in-
terpretation but its changes over time or across various reform scenarios can still be informa-
tive. It also produces indicators for expected returns of typical life careers. In addition,
PROST can be used to get to aggregate measures such as number of future contributors and
pensions (but not that of pensioners), revenues and expenditures of the system, financial bal-
ance of the system and the pension funds, and a special measure of sustainability, the balanc-
ing level of contributions or replacement rates. A further application of the model is to pro-
duce cumulative measures such as the implicit pension debt.
The model can assess both parametric and structural reforms, such as the introduction of indi-
vidual, funded retirement savings accounts. A shift from a defined benefit pay-as-you-go
scheme to one based on notional accounts can also be modelled including measurements of
transition costs.
In Malta, PROST has been catered for by the Ministry of Finance, Economy and Investments
and it has been updated in approximately every 3-4 years. It is mainly used in long-term
budgetary planning and the evaluation of pension reform options.

Netherlands
GAMMA
GAMMA (acronym for General Accounting Model with Maximizing Agents) is a dynamic
deterministic open-economy overlapping generations general equilibrium model of the Neth-
erlands, developed by the CPB (Netherlands Bureau for Economic Policy Analysis). The
model distinguishes markets for goods, capital, labour and a ‘market’ for income transfers
from the government to households. GAMMA identifies the following agents: households,
pension funds, the government, firms and the foreign sector. Households are divided up into
99 cohorts. GAMMA incorporates the economic behaviour of households, firms and pension
funds. Households decide on labour supply and private saving, firms decide on demand for
labour and capital, and pension funds decide on pension contributions and benefit levels.
Agents are rational and forward looking, and optimise in a consistent microeconomic frame-
work. GAMMA thus allows for welfare analysis of policy reforms. However, since perfect
labour and capital markets are assumed, the model is not equipped to describe short- and me-
dium-term dynamics.
GAMMA includes three separate sub-models: one for social security schemes (MOSI), one
for occupational pensions (EXPLOT) and one for occupational early retirement pensions
(PVK). As regards the system of public pensions, GAMMA relates the development of first
pillar pension expenditure to two factors: the productivity in the economy and the number of
people over the age of 65. Some age-specificity within this group is introduced to take ac-
count of the positive correlation between age and the proportion of single persons. For the
projection of the occupational pension pillar, the 700-plus pension funds in the Netherlands
have been assembled in a model of a single ’average’ pension fund. This average pension
fund offers a pre-funded average pay scheme, aiming at a replacement rate of 70 percent of
average pay and investing in a mixed portfolio of bonds and equity (50-50 percent). The pen-
sion model includes economic behaviour inasmuch as contribution rates and indexation fac-
tors are functions of the coverage ratio of pension funds, i.e. the amount of financial wealth in
terms of pension rights. The hypothetical model fund charges actuarial cost-effective contri-
bution rates (cost effectiveness being related to the entire pension fund) and indexes accrued
rights at combination of wage and price increase (full indexation above a funding ratio of 135

48
percent of the nominal liabilities). GAMMA has been used for the purposes of the AWG pro-
jections.

MICROS
The microsimulation model MICROS has been in use at the Ministry of Social Affairs and
Employment for several years. The model consists of a large micro database, a set of calcula-
tion rules (written in Fortran) and a large database with input variables.
The micro database consists of more than 60,000 households of whom data are available on
household composition, income and housing costs. The calculation rules represent, among
other things, the tax- and social security system. In this way, the income effects of a wide
range of policy measures in the field of income policy, including pension reforms, can be cal-
culated.
The model can simulate ageing in a static, as well as a dynamic way. In static ageing, the
situation of individual cases remains unchanged (only income levels change). The structure of
the population is revised by changing the weights of the cases. Dynamic simulation is the
simulation of events like birth, leaving school, finding a job or (early) retirement. Here, a
probability of each event is calculated for each case and subsequently is determined whether
the event occurs or not.
The dynamic ageing part models life events such as entering the labour market (with transi-
tion from education to work), marital status (marriage, divorce), childbirth, labour market
position (getting a job, changing working hours, losing job, losing working ability), leaving
the labour market (early retirement, retirement). The model also has a housing market module
and a migration module.
In practice, the Ministry of Social Affairs and Employment primarily focuses on the static
ageing version of the model. Important applications of dynamic microsimulations with static
ageing are developments in purchasing power, income inequality, poverty, and financial re-
turns on participation. The model adjusts its poverty and inequality measures to household
size and composition, and calculates various replacement rates, poverty measures (Leiden
poverty line, median poverty lines), and inequality indicators (Gini, Theil, decile ratios).

SADNAP
The SADNAP model (Social Affairs Department of the Netherlands Ageing and Pensions
Model) is a dynamic microsimulation model developed by the Department of Social Affairs
and Employment. The model is still under development. Currently, it is capable of projecting
state pension (AOW – Algemene Ouderdoms Wet) expenditures, but there are plans to up-
grade it to cover private pensions and allow for the analysis of income distribution (inequality
indicators, distribution of wealth among generations). The construction of a proper marriage
market module is also envisaged.
The model is based on two administrative datasets. Statistics Netherlands (CBS) supplied a
micro data file containing information on the number of acquired AOW entitlement years for
the population aged 15-64, along with the following individual characteristics: age, origin
(native, western non-native, non-western non-native), gender, marital status. The authority
responsible for the payment of state pensions, the Social Insurance Bank (SVB) supplied a
comprehensive dataset on all 2.6 million people receiving AOW in 2006, including data on
AOW entitlements, age, gender, country of residence, marital status, age of the partner and

49
the existence of a partner allowance. (This file also contains data on people receiving state
pension abroad.) The demographic assumptions for the projection were also provided by the
CBS. With regard to household structure, the tool can only distinguish between singles and
non-singles (through the use of a simple random procedure). Transitions between statuses are
simulated by the Monte-Carlo method.
The model is currently used for budgetary projections and for the assessment of various pol-
icy alternatives (e.g. removal of the partner allowance, increase of retirement age).

Austria
Austrian Applied Projection Models
Pension projections used to be made by the main consultative forum of policy discussions, the
Pension Advising Council (PAC). The PAC, composed of external experts and representatives
of the government and social stakeholders, presents an annual opinion paper to the govern-
ment. However, since the introduction of the sustainability factor in 2005, a calculation sig-
nalling unexpected changes, such as an unanticipated growth in life expectancy, a permanent
monitoring mechanism was launched in 2007. The mechanism does not automatically trigger
adjustment of contributions and benefits but mobilises an expert team to propose corrections
to such effect. The organisational background of the expert team is still to be clarified, in par-
ticular under the auspices of which government agency they work, or alternatively if they
work in isolation from the government.
The projection is based on two separate models reflecting the structure of the pension system,
which consists of separate schemes for the private sector and the civil service (as well as
smaller schemes for farmers and the self-employed, which are modelled according to the pri-
vate sector model). The two models are consolidated and cover the old age, disability, survi-
vors and early retirement functions.
The models calculate the number of new pensioners, the number of exits as well as future
expenditures.

Austrian microsimulation model (under construction)

The Austrian microsimulation model is under construction by an external contractor of the
Federal Ministry of Social Affairs and Consumer Protection. The aim of the Ministry is to
build the model to simulate the aggregate financial consequences of alternative policy propos-
als, and to be able to respond to ad hoc inquiries by various departments, political parties or
other stakeholders. The first module, the demography module, is expected to be completed by
2008.

Poland
FUS07
FUS07 model is a cohort based standard simulation model of the Social Insurance Institution,
performing deterministic actuarial calculations with an Excel – VB background. The model is
updated on an annual basis. It covers four social insurance schemes: old-age, disability, sick-
ness and accident insurance, however, it does not take into account the farmers’ scheme,
which is projected by Polmodel, a simulation model of the social policy budget. The tool dis-
tinguishes between different mortality rates of certain types of benefit recipients. Main out-

50
puts of the model include the standard fiscal indicators (expenditures, revenues, surplus/debt
of the Social Insurance Fund) and various standard ratios (e.g. benefit ratio). FUS07 (and ear-
lier versions) have served as a tool for the AWG pension projections.

Portugal
ModPensPor
ModpensPor, a cohort-based standard simulation model has been adapted by the Cabinet for
Strategy and Planning (GEP) of the Portuguese Ministry of Labour and Social Solidarity from
Spain’s Modpens model (developed by Fundación de Estudios de Economía Aplicada) with
the help of an external contractor in 1996. Besides pensions, the model can also project con-
tributions, unemployment, sick leave, and maternity benefits based on administrative data.
The model is programmed in Gauss Matrix Programming Language and updated on an annual
basis. ModPensPor is used for the assessment of financial sustainability of the social security
system, as well as for the AWG exercise. An area of further development could be the method
of determining the new flow of retirees. At present, it is performed in a simplified way, with-
out incorporating projected changes in the labour market. The proportion of a cohort becom-
ing eligible for pension is assumed to stay constant from the base year onwards, as is the case
with the average length of contributions (32 years). The wage history of new retirees is de-
rived from data on average wage growth since the 1960s.

Romania
No information was available for the research team on Romanian pension modelling.

Slovenia
SIOLG 1.0
The model SIOLG 1.0 is a dynamic overlapping-generations general equilibrium model of the
Slovenian economy, based on social accounting matrix (SAM) for the year 2000. The tool is
programmed in GAMS (general algebraic modelling system). The model has been developed
with the intention of analysing the sustainability of the Slovenian public finances as a whole,
though it can be used to analyse any sector of the economy (eg. pension, health care, long-
term care). The pension block enables the modelling of the first pillar of the Slovenian pen-
sion system, based on administrative data (Institute of Pension and Disability Insurance of
Slovenia). SIOLG 1.0 has been used in the AWG pension exercise.

Slovakia
MAJA
MAJA is a cohort model of the Ministry of Labour, Social Affairs and Family developed and
maintained in-house in cooperation with the ILO and the World Bank.

PROST (SK)
PROST (Pension Reform Options Simulation Toolkit) is a general-purpose model of the
World Bank. It is a combined semi-aggregate and typical agent model, in that it derives con-

51
clusions regarding macro expenditure and revenue flows by reweighing cohorts and genders
based on a demographic projection, but it is also capable of tackling as many as six typical
careers defined by the user. Consequently, it overcomes the usual problem of typical agent
models of producing aggregate results and at the same time it avoids some deficiencies of
cohort models such as remaining silent about replacement rates and similar life-cycle meas-
ures.
Accordingly, PROST produces the average entry pension over average gross wage type of
replacement rate. Indeed, this version of the replacement rate is not without problems of in-
terpretation but its changes over time or across various reform scenarios can still be informa-
tive. It also produces indicators for expected returns of typical life careers. In addition,
PROST can be used to get to aggregate measures such as number of future contributors and
pensions (but not that of pensioners), revenues and expenditures of the system, financial bal-
ance of the system and the pension funds, and a special measure of sustainability, the balanc-
ing level of contributions or replacement rates. A further application of the model is to pro-
duce cumulative measures such as the implicit pension debt.
The model can assess both parametric and structural reforms, such as the introduction of indi-
vidual, funded retirement savings accounts. A shift from a defined benefit pay-as-you-go
scheme to one based on notional accounts can also be modelled including measurements of
transition costs.
Slovakia has also used the PROST package to model pension contributions and expenditures.
Here it is operated by the Financial Policy Institute, a background institute of the Ministry of
Finance. For further details on the model see the above description.

Finland
Finnish Centre for Pensions Models
In Finland the primary centre of pension modelling is the Finnish Centre for Pensions.
The Centre’s standard model builds on the PTS standard simulation model for long-term
planning. The PTS is an actuarial valuation model, based on the standard cohort-based ap-
proach, which covers the entire earnings-related pension system. A separate model was cre-
ated for national pensions (minimum pensions), in which the number of eligible people is es-
timated using past information as a ratio of new pensioners to all non-pensioners.
The assumptions regarding the future age and sex composition of the population (including
mortality and fertility rates) are exogenous, however, the model is sophisticated relative to its
peers because for persons drawing an old-age pension a high pension is connected with a low
mortality risk when age and gender are standardised, and the mortality risk for disabled per-
sons is higher than the corresponding risk for the general population.
The model is not suited to produce micro level output, and results are annual pension expendi-
tures, contributions and accumulation of funds. Generation and gender-specific results are
also possible.
Standard modelling is occasionally extended by special purpose analyses, such as the ELPA-
analysis, the AK-analysis and the stochastic lifecycle model.
The ELPA-analysis aimed at calculating the effect of the pension reform on the pensions of
the 1998 retiree-cohort (about 40,000 people) based on administrative records of the Statistics
Finland.

52
The purpose of the AK-analysis is to project the future development of Finland’s earnings-
related scheme by extrapolating an administrative micro dataset, called the “lifetime wage
database” (AK). The AK is a 5000-person sample consisting of 400 to 1000 people from
every fifth birth-cohorts between 1905 and 1970. Data on earnings and detailed career infor-
mation are drawn from administrative records. The dataset covers the period of 1963 to 2005.
The time horizon of the model is 2045. The calculation concerns the development of new old-
age pensions, explicitly modelling the acquisition of pension accruals. While demography is
fully exogenous, an element of sophistication compared to other models is that the AK applies
different mortality rates depending on age at leaving school. Careers have been extrapolated
taking two distinct approaches: 1) historical data on how people’s earnings develop with age;
2) in a way consistent with the assumptions of the standard calculations. Distribution of pen-
sion wealth and future pension annuities payments are available, presented in the form of me-
dian, quartiles etc. Aggregate measures are under consideration; no formal calibration for ag-
gregation has yet been designed.
The stochastic life-cycle model was developed in 2005-2006. We received conflicting de-
scriptions whether or not this is a dynamic microsimulation model. The purpose of the devel-
opment is to evaluate the effect of the 2005 pension reform on retirement, employment and
pensions on the basis of financial incentives. The programming language was Gauss and
Stata. The underlying dataset consists of pension data from the Finnish Centre for Pensions
and data on employment from Statistics Finland. Data include wages and service years
throughout individuals’ employment histories. Age and gender are known for each individual.
The model describes labour supply decisions at the individual level. It includes detailed rules
of the Finnish statutory private-sector pension scheme and takes into account income taxation
and unemployment insurance. The parameters for the utility function are chosen so that as
with the pre-reform pension rules in force, individuals in the model behave similarly to the
way people behave in actual statistical data. The behavioural parameters are then fixed, but
the pension rules are changed and it is then observed how the individual’s behaviour changes.
The main outputs are standard replacement rates and tentative individual and aggregate return
measures.

Sweden
MiMESIS
MiMESIS has been used since 1973. It was built and maintained in-house by the social secu-
rity administration. It is frequently referred to in public discussions. The Swedish pension
reform was based on this model. The annual report of the Swedish pension system, the Or-
ange Report, was also built on it, aw were many governmental inquiries.
MiMESIS is a microsimulation model based on administrative individual data, which are not
matched with household figures or other surveys. It produces aggregate results on contribu-
tions and benefits for cohorts, gender and some other pre-defined groups. It explicitly models
labour income, old-age pensions, disability benefits and sickness benefits, but the main focus
is old-age pensions.
The model is mainly used to assess financial stability of the pension system under various
policy scenarios or demographic and macroeconomic conditions. It can also be used to calcu-
late replacement rates.

SESIM

53
SESIM (www.sesim.org) is a dynamic microsimulation model with dynamic ageing. It was
developed and has been maintained in-house by the Ministry of Finance. Program develop-
ment did not require specialised programming expertise; the developers had finance, econom-
ics and similar backgrounds. Accordingly, the program is written in Visual Basic, a well-
known software, and generates Excel and Access outputs. SESIM was built in 1997 as a tool
to calculate budgetary and distributional effects of student loans and grants. After 2000 it was
transformed to a pension model and since 2003 it has been used on a regular basis for that
purpose.
SESIM is based on the longitudinal panel of Statistics Sweden, called LINDA, which was
created from administrative registers, such as registers of the tax office and the social security
administration, with the start year of 1999. It contains about 380, 000 people, which is about
3.5 percent of the population. Combined with household members the sample grows to as
much as 786,000 people. LINDA is matched with the income distribution survey of Statistics
Sweden, named HEK, the GEOSWEDE spatial dataset (for data on regional mobility and ten-
ure), the household expenditure survey (for information on indirect taxes) and even the emi-
grants data of the National Social Insurance Board on people who do not reside in Sweden but
collected pension accruals there. Altogether, the input data of SESIM include information on
gender, age, marital status, household composition, education, income, wealth, taxes paid and
housing. Income data are derived from a wide range of income sources, such as labour in-
come, old-age pension, other pension benefits and even the value of in-kind public services
(education, health care). Data on households are also rather extensive. The modellers can
count on household size, household composition, age of members and household income.

United Kingdom
Pensim2
Pensim2 models demographic, education, and labour market processes such as, labour market
status and earnings, partnership and fertility, and mortality. For the elderly population sav-
ings, taxes, housing, institutional care, disability benefits and income related benefits are also
simulated. The lifetime accumulation of state pension contributions and private pension
membership are also modelled. Transition equations are estimated from either the LLMDB or
the BHPS. The model does not simulate behavioural responses.
Pensim2 output was one of sources used for policy formation by the UK Pensions Commis-
sion in its second report: A New Pension Settlement for the 21st Century. It is now established
as one of the major tools used by analysts in the Department for the development of State and
Private pension policies.

54
References

Baldini, M., Mazzaferro, C. and Morciano, M. (2008): Assessing the Implications of Long
Term Care Policies in Italy: a Microsimulation Approach. Department of Economics, Bologna
and Modena. Manuscript.
Belloni, M., Ferraresi, P. M., Maccheroni, C. (2003): Money’s worth measures of some re-
form scenarios for the Italian Social Security: Micro and macroeconomic effects using cohort
specific mortality tables. 4th International Research Conference on Social Security, Antwerp.
Berger, F., Islam, N., Liégeois, P. and Wagener, R. (2008): Cross-validating administrative
and survey data sets through the assessment of the 2001-2002 tax reform in Luxembourg.
CEPS/INSTEAD, Luxembourg. Manuscript.
CERP Activity Report 2000-2006.
http://servizi.econ.unito.it/cerp/site/research/activity_report_2000/at_download/attachField.
Decoster, A., De Swerdt, K., Orsini, K., Lefèbvre, M., Maréchal, C., Paszukiewicz, A.,
Perelman, S., Rombaut, K., Verbist, G. and Van Camp, G. (2008): Valorisation of the mi-
crosimulation model for social security MIMOSIS. Final report project AG/01/116. Leuven,
manuscript.
Decoster, A., De Swerdt, K., Orsini, K. and Van Camp, G. (2007): MIMOSIS: MIcrosimula-
tion MOdel for Belgian Social Insurance Systems - Modelling rules for household composi-
tion and family relationships. Social Security Working Paper.
Dekkers, G. J. M. (2007): Construction of a dynamic cross-sectional micro simulation model:
the state of affairs. Intermediate report 4.10, task 4.1. AIM Interim Research Report 4.10.
Brussels: manuscript.
Draper, D.A.G. and Armstrong, A.M. (eds.) (2007): GAMMA, a Simulation Model for Age-
ing, Pensions and Public Finances. CPB Document, No. 147. The Hague: Netherlands Bureau
for Economic Policy Analysis.
Dupont, G., Hagneré, C. and Touzé, V. (2003): Les modèles de microsimulation dynamique
dans l’analyse des réformes des systèmes de retraites: une tentative de bilan. ORCE Working
Papers 2003-10. Paris: OFCE Science Po.
Edin, P-A. and Fredriksson, P. (2000): LINDA − Longitudinal INdividual DAta for Sweden.
European Commission (2007): Pensions Schemes and Projection Models in EU-25 Member
States. European Economy Occasional Papers 35. Brussels: European Commission.
Flood, L., Jansson, F., Pettersson, Th, Pettersson, Th., Sundberg, O. and Westerberg, A.
(2005): SESIM III – A Swedish dynamic microsimulation model. www.sesim.org.
Frommert, D. and Heien, Th. (2006): Retirement pension provision schemes in Germany
1996 and 2005. Schmoller Jahrbuch 126. Duncker and Humblot, 329-336.
Gradus, R.H.J.M., van Sonsbeek, J-M., van der Werf, M. (2008): The SADNAP Model De-
partment of Social Affairs and Employment (Working Paper).
Harding, A., (ed.) (1996): Micro simulation and Public Policy. Amsterdam: Elsevier.
http://lesrapports.ladocumentationfrancaise.fr/BRP/064000302/0000.pdf
Kortmann, K. and Schatz, Ch. (2001): Retirement pension provision schemes in Germany
1996 – AVID 1996.

55
Maréchal, C., Perelman, S. and Van Camp, G. (2007): The distributive impact of alternative
Pension Welfare Adaptation reforms in Belgium
Mazzaferro, C. and Morciano, M. (2007): Pension reforms, educational choices and the long
term dynamic of employment in Italy. Department of Economics, Bologna and Modena. Pre-
liminary, manusript.
Mazzaferro, C. and Morciano, M. (2008): CAPP_DYN: A Dynamic Microsimulation. Model
for the Italian Social Security System
Schatz, Ch. and Merz, J. (2000): Die Rentenreform in der Diskussion Ein Mikrosimula-
tionsmodell für die Altersvorsorge in Deutschland (AVID-PRO). Manuscript:
http://mpra.ub.uni-muenchen.de/7328/.
Wiese, P. (2006): A Manual for The 2006 Version of Lithuania's Pension Simulation Model.
Manuscript.
www.fm.dk
www.etk.fi
www.ksz.fgov.be/nl/statistiques/stats_home.htmwww.ksz.fgov.be/fr/statistiques/stats_home.h
tmwww.sesim.org
Zaidi, A. and Rake, K. (2001): Dynamic Microsimulation Models: A Review and Some Les-
sons for SAGE. SAGE Discussion Papers 2. London: LSE ESRC.

56
Annex: The PENMICRO questionnaire

The PENMICRO questionnaire consists of questions regarding

- general information on the model
- micro datasets
- core model properties
- other model properties and
- output indicators.
Depending on the type of model (standard typical agent, standard cohort, static microsimula-
tion / dynamic microsimulation with static ageing and dynamic microsimulation with dynamic
ageing) we circulated different combinations of questions. From the question sets below the
following combinations were asked about
1. standard typical agent models:
G1-G9: General information on the model
A1-A5: Availability of micro data
TA1-TA7: Properties of models modelling careers of fictitious typical agents
R1-R11: Other questions on modelling
O1-012: Output indicators
2. standard cohort models:
G1-G9: General information on the model
A1-A5: Availability of micro data
C1-C6: Properties of standard cohort models
R1-R11: Other questions on modelling
O1-012: Output indicators
3. static microsimulation models / dynamic microsimulation models with static ageing
G1-G9: General information on the model
DP1-DP9: Micro dataset of the pension authorities (administrative data only)
DM1-DM8: Other datasets matched with pension data
MSA1-MSA6: Simulation properties of static microsimulation models or dynamic
microsimulation models with static ageing
R1-R10: Other questions on modelling
O1-012: Output indicators
4. dynamic microsimulation models with dynamic ageing
G1-G9: General information on the model
DP1-DP9: Micro dataset of the pension authorities (administrative data only)
DM1-DM8: Other datasets matched with pension data
MDA1-MDA8: Simulation properties of dynamic microsimulation models with dy-
namic ageing
R1-R10: Other questions on modelling
O1-012: Output indicators

The sets of questions are listed below.

57
General information
General information on the model
G1 responsible agency
G2 developer
1. imported from international agency, namely …
2. developed and maintained by external contractor for agency
3. developed by external contractor, maintained in-house
4. developed and maintained in-house

Is the external contractor academic or commercial?

G3 stage of development
G4 resources devoted to initial development
amount paid to external developer
did initial development include buying specific software? If yes, which one?
man-months devoted to data collection/preparation
man-months devoted to initial development (including debugging, calibration)
if own development: which kind of expertise was hired?
G5 maintenance
frequency and scope of updating
amount paid to external contractor for maintenance
man-months devoted to maintenance/updating
which kind of expertise was hired/employed?
further development
amount paid to external contractor for upgrading
man-months devoted to further development
G6 programming language
G7 application in policy formation or governance
G8 type of model
- macro
- standard (typical agent)
- standard (cohort, aggregate)
- dynamic MS with static ageing
- dynamic MS with dynamic ageing
G9 Which schemes / benefits are considered in the model?

58
Micro datasets
Availability of micro data
A1 Are there micro datasets available containing information on
education history
employment history
wage history
contribution history
accumulated accruals
health conditions, public health expenses by individual
household panel (income, consumption)
A2 Are some of these datasets matched? If yes, which of them?
A3 Are there technical or legal constraints of matching micro datasets?
technical
legal
A4 Would there be possible to match entries of a survey dataset with administrative data? If yes, which ones?
A5 Would there be possible to match entries of an administrative dataset with other administrative data? If yes,
which ones?

59
Micro dataset of the pension authorities (administrative data only)
DP1 name of the dataset(s) proceeded by the model; size of samples
DP2 Demography
Which of the listed characteristics are recorded?
- ethnic/racial origin
- migrant status
- gender
- age
- education
- marital status
- mortality by scheme (old-age, disability, etc)
- region
DP3 employment
what kind of observed data can be gained on the employment history of the individual?
How far do data go back with retrospective information on employment?
DP4 wages, income
What kind of observed data on current and retrospective wages does the dataset contain?
How far do data go back with retrospective information on wages?
What kind of observed data is available on other sources of income?
DP5 separate scheme information
Is there information available on accruals in other mandatory schemes such as insurance
schemes or second- and third-pillar retirement schemes?
Can these datasets be matched with data on the first-pillar?
DP6 matching data
Can the benefit of observed pensioners be matched with previous data on contributions?
What other administrative datasets (such as health conditions, employment, tax declara-
tion, wealth/property register) can administrative pension data be matched with?
- if none:
• are such datasets nonexistent?
• are they not matchable?
• are there legal constraints to match them?
DP7 individual vsersus household data
Matching strategy: does it start with pension authority data and match it with survey data,
or the other way around?
Can the individual dataset be matched with administrative data on other household mem-
bers?
Can the individual dataset be matched with household survey?
DP8 Are data on past pension accruals available (for which period / scheme)?
DP9 Does the set contain accumulated accrual / pension wealth data?

60
Other datasets matched with pension data
DM1 Administrative or survey or both; size of samples? Name of the datasets proceeded by the model.
DM2 are individual data matched with household data; if yes, what is the method of matching?
DM3 Are separate micro datasets matched; if yes, what is the method of matching?
DM4 Which of the listed individual characteristics are recorded in matched data?
- ethnic/racial origin
- migrant status
- gender
- age
- education
- marital status
DM5 What are the income components covered?
- labour income
- old-age benefit
- survivor benefit
- disability benefit
- social assistance
- value of in-kind services
- other: ...
DM6 Which of the listed household characteristics are recorded?
- size
- age of members
- combined income (including income definition)
DM7 Time in data
- cross section
- repeated cross sections
- retrospective data from cross section set
- panel
DM8 What kind of problems did you face and solutions did you find in the matching procedure that could be
relevant for modellers in other countries?

61
Core model and simulation properties
Properties of models modelling careers of fictitious typical agents
TA1 What are the characteristics of typical agents whose life path is modelled?
TA2 How are typical agents selected and how are their life-path characteristics established?
TA3 Do data allow extending the characterisation of typical agents with other features?
examples:
- ethnic/racial origin
- migrant status
- education
- occupation
- marital status
- health conditions
TA4 What do life-paths of typical agents look like?
TA5 What are the core scenarios or assumptions of demographic, labour market and other relevant develop-
ments in drawing the life-paths of typical agents?
TA6 How is consistency between core scenarios and/or assumptions established or maintained?
TA7 What is the procedure, if any, applied to derive aggregate results from the life-paths of typical agents?

62
Properties of cohort models
C1 What are the characteristics/subgroups considered in projecting the career of a cohort?
examples:
- age
- gender
- level of education
- labour market position
- occupation
- wage
- migrant status
- ethnic/racial origin
C2 What are the classes of the classification among the dimensions reported in C1?
age
level of education
labour market position
wage
ethnic/racial origin
others: …
C3 What are the assumptions regarding future developments of the following characteristics/subgroups?
population, age composition
mortality, fertility
ethnic/racial composition
migration
family composition, marital status
occupation
health status
C4 What does the reweighting of subgroups, if any, as an imitation of future developments looks like?
C5 What are the assumptions regarding the following macroeconomic variables?
inflation
interest rate (short and long term)
GDP
wage
real rate of return
employment
real income growth
C6 What does the mechanism ensuring consistency among the assumptions regarding various characteristics /
macroeconomic variables look like?

63
Simulation properties of static microsimulation models or dynamic microsimulation
models with static ageing
MSA1 What are the dimensions in the multidimensional re-weighting matrix
MSA2 What are the classes of each dimension? (including the option ‘continuous’)
ethnic/racial origin
migrant status
education
marital status
labour market status
occupation
health conditions
other
MSA3 What are the other components of the re-weighting matrix?
MSA4 What is the data source of the various dimensions of the re-weighting matrix (possibly specified to
gender) and of the projected aggregate controls in the updating process?
population, age composition
gender
ethnic/racial origin
migration
family composition, marital status
wage
real rate of return
employment
occupation
inflation
interest rate (short and long term)
GDP
real income growth
MSA5 If re-weighting is carried forward serially in various dimensions: what mechanism (if any) keeps the
separate re-weighting processes consistent?
examples:
- education and employment
- education and occupation
- education and wages/contributions
- marital status and fertility
- disability/health and employment
MSA6 What kind of problems did you face and solutions did you find in the modelling procedure that could be
relevant for modellers in other countries?

64
Simulation properties of dynamic microsimulation models with dynamic ageing
MDA1 cross-sectional vs cohort/longitudinal
MDA2 What are the life events modelled: what are the stata in these processes?
education
marital status
birth, adoption
labour market position
occupation
retirement
health
other: …
MDA3 Which processes are simulated via endogenous transition probabilities, produced by logits or probits?
MDA4 What is the source of the exogenous transition probabilities in these processes?
survival
education
marital status
birth, adoption
labour market position
occupation
retirement
migration
health
other: …
MDA5 What kinds of empirical models are used to simulate the values attached to the individual in these proc-
esses?
survival
education
marital status
birth, adoption
labour market position
occupation
wages
retirement
migration
health
other: …
MDA6 What is the order of modules in the recursive calculation?
MDA7 If transition probabilities are calculated separately: what mechanisms keep consistency among them?
examples:
- education and employment
- education and occupation
- education and wages/contributions
- marital status and fertility
- disability/health and employment
- other: ...
MDA8 What kind of problems did you face and solutions did you find in the modelling procedure that could be
relevant for modellers in other countries?

65
Other questions on modelling
Other questions on modelling
R1 Does the model apply mortality rates differentiated across various stata (disability, marital status, ethnic
origin, migrants, type of pension scheme)?
R2 Are migrant workers treated separately?
R3 Can the model address the level of informality in the economy (including measuring the covered wage
bill)?
R4 Is the acquisition of pension accruals explicitly modelled?
R5 How are minimum pension arrangements modelled? Are there any necessary simplifying assumptions
made? If so, what?
R6 How does the model deal with portability of accruals across national pension schemes?
R71 Is the share of part-time work projected?
R72 How does the model simulate taxes (other than social security contributions)?
R8 Can effects of pension reforms on the social assistance budget modelled?
R9 Does the model allow for parallel work and retirement?
R10 What kinds of empirical models are used to derive behavioural responses in these areas:
fertility
labour supply
savings
R111 What kind of retrospective data regarding contributory history or accumulated accruals does the model
use?
1
Only standard models.
2
Only microsimulation models.

66
Output indicators
Output indicators
O1 What kind of replacement rates can the model produce?
O22 If the model contains household income data: what is the equivalence scale applied?
O3 What kind of poverty measures can the model produce?
O4 What kind of income inequality measures can the model produce?
O5 What kind of non-pecuniary measures of poverty and inequality can the model produce?
such as measures of non-monetary deprivation, in particular, in housing
O6 Does the model produce cumulative values such as pension wealth or related measures?
O7 What kind of individual return measures does the model produce?
implicit rate of return
implicit tax rate
payback period
other: ...
O8 What kind of aggregate return measures does the model produce?
implicit rate of return
implicit tax rate
other: ...
O9 What kind of sustainability measures does the model produce?
- number of future contributors and beneficiaries
- number of future minimum pension recipients
- future revenues and expenditures
- implicit pension debt, net present value
- other: ...
O10 What kind of intra- or intergenerational redistribution measures are produced by the model?
O11 Can the model decompose tendencies in aggregate expenditures/ revenues and if yes into what components?
O12 What kind of output or indicators other than mentioned in the previous questions can the model produce?
2
Only microsimulation models.

67