The dynamic cross-sectional microsimulation model MOSART

MOSART is a dynamic microsimulation model for education, labour supply and pensions. The core idea is to draw a sample of the population, and then simulate the further life course for each person. According to Orcutt (1957) and Orcutt et al. (1986) microsimulation is a useful method when behaviour is more easily modelled at the individual level, and then aggregated to totals, rather than modelling relations between aggregated variables. This is often the case when behaviour is non-linear and the population heterogenous, blurring the relation between micro and macro. Analyses of tax- and pension-systems are good examples.

MOSART is based on administrative data for the entire Norwegian population back to the 1960s and simulates the further life course for each person. The simulation is carried out by drawing if each person each year makes certain transitions from one state to another with transition probabilities depending on each person’s characteristics. These probabilities are estimated from observed transitions in a recent period. For years with actual observations simulation and transition probabilities are aligned to these. As common in several other dynamic microsimulation models (cf. O’Donoghue and Dekkers, 2018), events like births, deaths, migration, marriages, divorces, educational activities, labour force participation, retirement, income and wealth are included in the simulation. New-borns and new immigrants are added to the initial population each year. The earlier, but also the most fundamental, parts of the model are documented by Fredriksen (1998).

Securing the necessary financing to develop and maintain large scale microsimulation models such as MOSART can be a challenge in the long run. Like models in France, Canada and US, MOSART is situated in a statistical office and thus benefits from easy access to microdata. In addition, the MOSART project is fortunate to be embedded within a strong tradition of model-based policy analysis. Laying the groundwork for policymaking in Norway has for more than half a century been a collaboration between The Ministry of Finance (and other ministries) and Statistics Norway, with theoretical support from the academic sector, especially the University of Oslo.¹ It is still the case today that Statistics Norway operates a wide selection of planning models actively used by the government, including both static and dynamic microsimulation models, macro models and environmental models.

The main purpose of MOSART is thus to facilitate government planning. This has resulted in emphasis being given to developing and maintaining a large, detailed and up to date database, and to doing simulations where current behaviour (often in the form of constant transition probabilities) continues indefinitely into the future. This constant behaviour assumption is unrealistic, but uncontroversial and easy, at least in principle, to understand. This facilitates planning because it precludes difficult discussions of how to model future behaviour (and governments might have their own thoughts on behaviour), while providing a consistent framework within which to discuss policy. Since MOSART has a central place in policy discussions it is also occasionally used by other public institutions, private organisations and the media. Even though policy analysis has priority, academic work is important in maintaining the quality of the model. An example of this is the work on mortality and life expectancy documented in Brinch et al. (2018).

Since the end of the 1990s MOSART has been used extensively in analysing the need for a pension reform, in planning the reform enacted in 2011, and in the evaluation of its effects. Based on similar ideas found in Dekkers et al. (2018) in their analyses for Belgium, Sweden and Italy, MOSART has been used for both (i) assessing the budgetary impact of ageing and different designs of the reform and (ii) assessing pension adequacy/distributional questions. While budgetary impacts from ageing have been analysed by microsimulation models in several countries, see Dekkers and Van den Bosch (2016), MOSART has also been extensively used in distributional analysis. In contrast to most other countries, MOSART was also extensively used in developing the pension reform.

In a dynamic microsimulation model, pension benefits are calculated from labour market earnings and other relevant characteristics according to an accurate description of the pension system. At present the pensions covered by MOSART include old-age pensions, disability pensions and surviving spouse pensions in the National Insurance Scheme (NIS), early retirement benefits in both the private and the public sector, and occupational pensions in the public sector. In addition to the analyses of effects from the pension reform, the model has also been used regularly for projecting the labour force by level and field of education.

In the next sections we will present and discuss simulation techniques, technical platform and data used. The main data input is based on administrative registers and contain information about the entire population. Thereafter we describe and discuss the design of the main blocks of the model regarding demographics, education, retirement, labour market, earnings, pensions and fiscal effects. In addition to giving an account of how MOSART is used in analysing fiscal effects of the Norwegian pension reform we discuss different approaches used by the model to analyse distributional effects. Finally, we present how we use the model to calculate accrued to date entitlements.

Dynamic microsimulation covers a large variety of life cycle models. Further development after the pioneering work by Orcutt (1957) is discussed in among others O’Donoghue (2001), Wolfson (2009), Li and O’Donoghue (2013), O’Donoghue (2014), Dekkers and Van den Bosch (2016) and O’Donoghue and Dekkers (2018). Dynamic microsimulation models can be grouped in two different types: (i) the longitudinal type, and (ii) the cross-sectional type. The difference between the two groups depends on the order in which individuals are simulated over time. Simulation in MOSART may be characterized as cross-sectional because all aspects of the population in one year are simulated before the model starts simulating the next year. By contrast, longitudinal simulation models simulate one individual for all years between one and T before simulating the next individual.

A cross-sectional model is advantageous when simulating the distributional effects of a change in the pension system because it allows for interaction between individuals. Dekkers and Van den Bosch (2016) conclude that cross-sectional models are therefore more suitable than longitudinal for the simulation of household income, inequality, poverty risks and adequacy of social security and pension benefits. Longitudinal models are more suited to analysing the budgetary impacts of ageing.

Dekkers and Van den Bosch (2016) also point out that cross-sectional microsimulation models are more common than longitudinal. The models used in France, Sweden, Belgium and Luxembourg, Hungary, UK and US are of the cross-sectional type. In most of these countries the models are also used to simulate budgetary costs of ageing, but in Belgium and Luxembourg these costs are simulated by a semi-aggregate model. In Norway budgetary costs connected to pensions are simulated using the MOSART-model, while budgetary costs related to an increasing amount of health and care services, due to an ageing population, are simulated by a semi-aggregate model.

A further common characteristic for MOSART and the cross-sectional models used in the other countries mentioned above is the fact that they are based on discrete time, cf. Li et al. (2014). The combination of discrete time and cross-sectional simulation has the advantage that it allows to use exogenous constraints on the number of events or objects within each time unit. External constraints may either be based on actual data for historical years, or on time series from e.g. a macroeconomic model. The demographic projections in MOSART are aligned to the results from Statistics Norway’s official population projections. The system for adjustments in the model is based on a (hypothetical) logistic representation of each probability p_i for a person ‘i’ (or a specific group):

Here X_i is the vector of covariates for person (group) i, and β denotes the corresponding coefficients. The adjustment term δ is common to all (or a specific group) and is adjusted so that the sum of the probabilities gets (close enough) to its constraint.

An obvious weakness, however, is that this method does not allow for competing risks. A significant weakness in MOSART is that the year was chosen as the basic time unit. Most pension benefits do not follow the calendar year but occur within a year. With todays’ access to data and far larger computer capacity compared to the beginning of the 1990s, when MOSART was established, we probably would have chosen month instead of year as the basic time unit. With this choice, several of the existing weakness in aggregating over the year would have been eliminated.

A discrete-time model requires that each event is simulated sequentially. As pointed out by Klevmarken (2008) it is unavoidable that the modeller therefore must choose an order by which these various characteristics can be changed within each period. The order of the processes in recursive simulation is important since earlier events may affect later events in the same period. The modelling in MOSART also follows the classical structure of such models with ‘recursive blocks’ that simplified may be presented as (i) the demographic block, (ii) the education block, (iii) the retirement block, (iv) the labour market block, (v) the earnings and income block, (vi) the pension benefits block and (vii) the fiscal block.

Like the cross-sectional dynamic microsimulation models in the other countries mentioned above the key methodology used in MOSART is stochastic simulation. As common in the other models, simulations are based on the confrontation of the deterministic part of a logistic regression model with a logistic random variable. The simulation will normally also use variance reduction methods. The simplest (and most common) variant is binomial events, where the expected number of events is equal to the sum of transition probabilities. By adding up these probabilities as each individual is handled, an event occurs every time the sum passes an integer value. The number of events will be constrained to its expectation value, and each individual will with some precautionary steps have a probability of experiencing an event that is equal to its transition probability. Variance reduction methods have a decisive effect on clarity, particularly that even small deviations cannot be explained away as mere simulation noise.

Simulation in MOSART starts in 1966, the year prior to the foundation of the National Insurance Scheme, including the start of most of the data and the earning of pension entitlements. The model runs through all historical years exactly as the simulation does, including the recursive part, except that the drawing of each characteristic is replaced with an imputation of the actual historical value. The reason for this apparent complication is the pension reform. Old age pension benefits will by this procedure depend on time of withdrawal and possible earning of pension entitlements after withdrawal. Full access to all modules for historical years is also an upside. We will for this reason have a close to seamless transition from history to simulation, with access to data and tables for historical years and improved possibilities for testing the model. As mentioned above, alignment methods using binary techniques to constrain model output to conform closely to externally derived macro-data are used to a large extent in MOSART.

In analyses of the pension system, pension benefits may be calculated with different set of rules within one simulation, and this aspect is generalised to all pension and tax rules. The simulation has a ‘pension system one’ that interacts with the simulation. An additional pension system may be requested and will calculate pension benefits from the same life courses with one or several changes in pension rules, but without any effect on the simulation. Different ‘rules’ may include withdrawal of pension at other ages and exclusion/increase of pension entitling income at specified ages. Multiple calculations of pension benefits remove almost all remaining simulation noise from the effect of changing a pension rule, making even small changes appear clear in magnitude and timing.

MOSART utilizes administrative data registers available in Statistics Norway, covering the entire population back to the 1960s and 1970s. These data form the initial population of the model, but they are also used to estimate parameters and for testing. Maintenance and development of data have two distinctive steps. The first step is collecting new data and organizing them into what we refer to as the adapted source data, maintaining the original information and including variables that may come into use in the near future. The second step produces the initial population, a version where missing and/or inconsistent data are corrected, with a format streamlined for reading by the simulation model.

The modelling of demographic events and the underlying demographic assumptions in MOSART are to a large degree based on the official population projections from Statistics Norway, see Leknes et al. (2018). The aggregate development in the main demographic components like fertility, mortality, in- and outmigration and domestic migration is thus aligned to fit the observed development combined with results that derived from the official population projections. However, it is quite easy to align the simulations to other aggregate assumptions, eg, population projections from Eurostat. More detailed fertility and mortality simulations in MOSART are used to decompose the aligned aggregates. Most importantly, MOSART also includes cohabitation and marriage as covariates, because fertility and other events depend on cohabitation/marital status, and this is also the case for some of the rules in the pension system.

4.1 Fertility

Births are estimated for women depending on age, number of children, age of the youngest child, cohabitation and educational activity. Each time a birth occurs, a child is added to the model population. Adjustments to make the simulated total number of births equal to observed levels or projected levels from the population projections, is roughly the same as adjusting individual fertility rates against the periodic total fertility rate (TFR). For each cohort TFR can be estimated as the average number of children per female in the cohort at an age of 49 years. Periodic TFRs are estimated as the sum of each period’s age-specific fertility rates.

The development in TFRs over the last 100 years is earlier documented by Brunborg and Mamelund (1994). Observations from the following years are added in Figure 1. Large variations in the annual number of births and thereby TFRs during the first half of the 1900s are still important for understanding the dynamics of the number of elderlies and projections of pension expenditures with the MOSART-model. The most important aspect is the low number of births between the two World Wars due to the economic depression. An increase in fertility from 1935–1946 measured by periodic TFR can probably be explained by postponing births until the economic depression and the following war was over. The large birth cohorts from 1946 until about 1970 replacing the small cohorts born between the wars is in combination with higher life-expectancy the main reason behind the expected large increase in Norwegian old-age pension expenditures from 2015–2040.

Figure 1

Download asset Open asset

The long-term falling trend in the cohort specific fertility rates since the peak of 2.6 for women born in 1935 is mainly caused by fewer women giving birth to three children or more. Falling fertility is a part of the so-called demographic transition most developed countries have experienced with first lower mortality and then lower fertility. Therefore, the periodic TFR dropped significantly up to 1977. After staying rather constant from 1977–2009 periodic TFR has dropped further in the last years and reached 1.62 in 2018, the lowest periodic TFR observed ever. In the medium alternative of Statistics Norway’s population projection from 2018 TFR is assumed to increase in the coming years to a long-term level of 1.76 from about 2030.

4.3 Migration

In- and outmigration by age, gender and three regions of origin from the population projections are used as exogenous input to the MOSART-model. In the model for migration in the population projections documented in Cappelen et al. (2015) in-migration depends on incomes per habitant in Norway compared to the three regions of origin, difference in the rate unemployment, the number of persons from each region of origin who have immigrated before and growth in population in the three regions:

1. Western Europe, Canada, US, Australia and New Zealand

2. Countries from Eastern Europe members in the European Union

3. Rest of the World

Out-migration is modelled by probabilities from the past ten years dependent on region of origin (Norway for non-immigrants), gender, age and period of residence.

In MOSART education is an important covariate for labour-market participation and incomes. Poor registration of level and field of education for immigrants has been a major weakness in projecting immigrants’ performance in the labour market. Several of them get the characteristics ‘unknown education’ implying lower labour market participation rates and higher risk of disability than for the average population. A large survey on immigrants’ educational background was conducted in 2011/2012, decreasing the share with unknown educational background substantially (Steinkellner and Holseter, 2013).

Immigration to Norway reached a peak in 2011 and 2012. In the following years immigration has decreased while out-migration has increased. Thus, net-immigration in 2017 was only one half of net-immigration in 2012. The fall in immigration in this period applies to almost every country of origin except Syria. By the three regions of origin the decrease in immigration is significant for both group 1 and 2, while the sharp increase in the number of immigrants from Syria has caused higher immigration from group three as a total.

According to the assumptions from Statistics Norway’s population projections from 2018, immigration from region one is expected to increase slightly in the first years due to positive business cycles in Norway. From 2021 and onwards immigration from this region is expected to be slightly reduced in the medium alternative due to lower growth in incomes in Norway than in the other countries of region 1. By the same reason immigration from region two is also expected to decrease. Because a large part of the immigrants from region three are refugees, there is lot of uncertainty in projecting future immigration for this group. With the medium alternative of the projections we assume immigration for this group to increase slightly Figure 2.

Figure 2

Download asset Open asset

Projections of population by main age groups based on the main alternative from Statistics Norway’s population for selected years are presented in Table 1. Total population is projected to increase by about 600 thousand from 2018–2035 and by 1.2 million from 2018–2060. Because of increasing life-expectancy and large cohorts born after Wold War II, the number of persons in age 80+ is expected to almost double from 2018–2035 and triple to 2060. A rather strong growth is also projected for persons in age 67–79 while population growth in age groups 0–19 and 20–66 is rather modest.

Table 1

Total	2018	2035	2060
	5332	5938	6530
0–19 years	1 258	1 277	1 344
20–66 years	3269	3455	3600
67–79 years	580	755	885
80+ years	226	425	702

1. Source, Statistics Norway.

Education is a characteristic that is important for a series of outcomes, such as attachment to the labor market, pensions and mortality. Educational background is not always implemented in microsimulation models, but some exceptions exist. Most commonly known are the Danish microsimulation project DREAM (Rasmussen and Stephensen, 2014) and CEPAM-mic used for EU-28 (Marois et al., 2017), although the disaggregation of education is less detailed than in MOSART. In MOSART, we define levels and fields of education, and project 28 different educational categories. The levels used are consistent with the definitions denoted ISCED 2011 from Unesco Institute for statistics, 2012.

Lower secondary education is comprised of compulsory education in Norway, and currently it lasts from age 6–16. Upper secondary education includes all who have completed three or four subsequent years in schooling, granting a degree at upper secondary level. Post-secondary, non-tertiary education is also included in this category. Bachelor level education includes education in universities or university colleges, and requires having completed two subsequent years, and most often three. Education at master’s level consists of two more years in advanced programs, and education at Ph.D.-level is also included.

Assigning educational characteristics to those in the future population who have not yet completed their education is done by simulating educational transitions. Assigning educational characteristics is an integrated part of the simulation model in MOSART, as opposed to e.g., in the Danish microsimulation model DREAM (Rasmussen and Stephensen, 2014). Start, level and field of education, completion of ongoing educational activity and, conditional on not completing, continue in education the next year are the main events used to assign educational background to the future population. For each year, all individuals have the possibility to make this series of transitions and their educational characteristics are updated upon completion.

The probabilities of the events described above are estimated using a series of Logit estimations, controlling for individual characteristics and educational history. All event probabilities are estimated separately according to gender allowing for individual characteristics to affect events differently for men and women.

Many of the migrants in the data are not registered with any educational attainment during the first years after arrival. Migrants who have education at upper secondary or tertiary level from their home country can choose to report their education, and if being approved by the Norwegian Agency for Quality Assurance in Education (NOKUT), it is registered in the educational register. To improve the data for migrants with unreported education, there was a survey in 2011, which has succeeded in improving the data quality substantially. However, 24% of the immigrant population were still registered without any educational attainment after the survey.

In the recent simulation, educational transition probabilities have been assumed to remain at the same level as the mean between 2009 and 2014, conditional on the educational and individual characteristics in the population. Over the last decades, there have been strong trends in favor of starting in higher education. However, when conditioning on individual characteristics and earlier educational attributes, there are no indications of such trends. Moreover, realized educational transitions are a combination of the behavior of individuals making choices and institutions offering places in education. To disentangle the two is challenging. Any trends in observed educational transitions could therefore not be interpreted as trends in behavior. Moreover, educational enrolment varies over the business cycle, and using an average over several years makes it possible to make projections that are neutral to business cycle conditions.

Figure 3 shows the development in education level from 1990–2040. High enrolment to tertiary education during the 1990s led to a substantial increase in the share of the population with higher education. The number of those with upper secondary education has also increased during the last 20 years, but with high propensities to study longer, the group is in fact diminishing over the next years. From around 2035, the group with education at bachelor level will become the largest group instead of upper secondary education. Persons with lower secondary education has decreased, and with constant educational propensities, they will continue to do so.

Figure 3

Download asset Open asset

Based on former labour market participation and incomes, a set of transition rules and information about other individual characteristics MOSART uses a detailed set of pension rules for pension arrangements included to calculate pension benefits for every individual in the population. This information may thereafter be used to calculate consequences for pension expenditures and distribution of income from different designs of the system. Since the end of the 1990s the main use of the MOSART-model has been to analyse direct fiscal effects as well as distributional effects from different alternative designs of main components in the Norwegian pension system.

Projections with the model are naturally dependent on the assumptions used, and sensitivity for some of the most important ones is briefly discussed by Fredriksen et al. (2017). Under the old pension system both expenditures and the distribution of benefits were highly affected by assumptions regarding mortality. Under the new system effects on expenditures from different assumptions regarding mortality and withdrawal behaviour are highly neutralized. For special groups like the disabled, those receiving widow/widower benefits and migrants, relatively large uncertainty regarding labour market participation and incomes causes a large uncertainty in projections of their pension benefits.

Figure 5 compares two simulations of pension benefits, ‘Mosart’ and ‘Sim1980’, with the observed benefits reported by the Tax Authorities, labelled as ‘Reported’. The figure shows the level of pension benefits sorted by increasing values. The simulation ‘Mosart’ simulates benefits keeping other variables equal to their observed values. It shows how well the pension module works in isolation. As can be seen from the figure it tracks very well the observed data. The simulation ‘Sim1980’ shows the simulated benefit in 2017 resulting from an historical simulation starting with the observed population in 1980. For each year the model is calibrated against observed aggregate statistics. We thereby focus on the quality of modelling the distribution of labour incomes and pension benefits. This simulation also tracks fairly well, even though, as the figure shows, too many men receive low benefits (and thereby too few receive high benefits). This may be caused by problems simulating the benefits of immigrants with a short period in Norway, because their labour incomes (which determine pension benefits) are simulated at too low a level.

Figure 5

Download asset Open asset

The main emphasis in the analysis of the Norwegian pension reform has been on the old-age part of the National Insurance System (NIS) that was established in 1967. NIS constitutes the first pillar of the Norwegian pension system and is by far the most important component for old-age pensions in Norway. As discussed by Christensen et al. (2012), and not influenced by the reform of the system in 2011, NIS is an integrated part of the Central Government budget and financed PAYG. Yearly expenditures do not have to be balanced by specific contributions, and pension expenditures are financed by general tax revenues. From 1973 until the reform in 2011, old-age pensions in NIS could be claimed with defined benefits (DB) from age 67 and were earnings-tested until age 70, until a stepwise repeal of this test was imposed between 2008 and 2010. Because of the early retirement schemes introduced in 1989 covering about 60 per cent of the employees in the private sector and all employees in the public sector, it was possible for employees belonging to these schemes to retire from the age of 62 without any consequences for benefits from NIS after the age of 67. Towards the end of the 1990s it was evident that growing longevity and large cohorts born after World War II would cause expenditures in this DB-system to increase substantially in the coming decades.

The work on reforming the system started by appointing a Pension Commission in 2001. In its report from 2004 (NOU, 2004/1) the Commission’s main suggestion was to reform the Norwegian old-age pension system including the early retirement schemes toward a nonfinancial defined contribution (NDC) scheme like the one already implemented in Sweden. Two agreements in the Norwegian Parliament in 2005 and 2007 indicated that a broad political majority supported a reform along the suggested lines, with some adjustments. Based on a tripartite agreement between the Government, the trade unions and the Employers’ Confederation the early retirement scheme in the private sector was incorporated in the new system from 2011. No corresponding agreement was reached between the Government and the trade unions in the public sector. The early retirement scheme and occupational old-age pensions in the Government sector were therefore only partly adapted to the new system leaving the early retirement scheme unaffected. However, by new negotiations an agreement to change these schemes corresponding to the system in the private sector was reached in 2018 with major effect from 2025 when those born in 1963 reach the age of 62.

Under the new system, old-age pensions may be drawn partly or completely between the age of 62 and 75, with actuarial adjustment and without any earnings test. From accumulated entitlements at retirement age, annual pension benefits for a cohort retiring are calculated by divisors reflecting remaining life expectancy at that age. Early retirement leads to lower annual benefits because accumulated entitlements must be divided by more years. This is also the case when life expectancy increases for a given retirement age. Lower benefits when life expectancy increases may be counteracted by postponing retirement. The connection between pension entitlements and lifetime labour incomes is also closer in the new system than in the old.

The MOSART model has been used to analyse effects on labour supply and old-age pension expenditures from the pension reform during the whole reform process as well as in post reform evaluations. Given the pre-reform uncertainty regarding labour-supply effects, assumptions were made based on the incentive structure in combination with possible effects from these changes discussed in the economic literature. In the post reform evaluation direct effects on participation rates for different groups according to analyses by Hernæs et al. (2016) are considered.

The full effect measured in bill. NOK in fixed 2017 wage amounts from the reform of the system for old-age expenditures in NIS compared to continuing the old system is shown in Figure 6. Counteracting further growth in life-expectancy through life-expectancy adjustment is the main tightening component in the new system. In 2050 old-age pension expenditures in NIS are simulated to be reduced by almost 20 per cent because of the reform while the expenditures in 2100 is estimated to have declined by more than 30 per cent. However, the reform does not counteract the effect from relatively large cohorts born after World War II replacing small cohort born in the former decades. Even with the pension reform old-age pension expenditures are estimated to more than double from 2010–2050. The increase beyond 2050 is weaker because of more stable sized cohorts of pensioners.

Figure 6

Download asset Open asset

Figure 7 shows the effects on pension expenditures from alternative designs of the old-age occupational pension system in the public sector. The full reform effect based on the agreement from 2018 is compared with a prolonging of the old system existing before 2011. The main tightening of occupational old-age pension expenditures in the public sector compared to the no-reform scenario took place by the general pension reform in 2011. The tightening was caused by lower indexing of pensions in payments than general wage growth as well as life-expectancy adjustments. The first effect from the 2018-agreement starting to work in 2025 will be a postponement of expenditures for early retirement benefits as this component is converted into a life-long supplement.

Figure 7

Download asset Open asset

The tightening of the old-age occupational pension system in the public sector will be relatively stronger than the tightening of NIS. In NIS some of the tightening caused by life-expectancy adjustment is somewhat compensated by more generous rules for accumulation of entitlements. In 2050 expenditures for old-age occupational pensions in the public sector are projected to be reduced by 31 per cent compared to a 21 per cent tightening of the expenditures from NIS. A significant increase in the number of employees in the public sector during the past decades will even with the full-reform system cause expenditures for old-age occupational pensions in the public sector to almost double from 2010–2050.

Adequacy and distributional effects played a prominent role in discussions prior to the Norwegian pension reform in 2011. As pointed out by Dekkers and Van den Bosch (2016), a model like MOSART, belonging to dynamic-ageing microsimulation models of the cross-sectional type, is well suited for these kinds of analyses. By using MOSART it is possible with a far more accurate description of distributional consequences of the 2011 reform than a traditional simplified approach restricted to the calculation of replacement rates given assumptions such as fixed level of income for 40 years. Simplified calculations of replacement rates presented in Christensen et al. (2012), show that persons with somewhat above-average income will gain from the change in accumulation of entitlements because of the Norwegian pension reform before taking life expectancy adjustment and lower indexation of benefits in payments into account.

In a first round of analyses of distributional effects from the Norwegian pension reform using the MOSART model also presented in Christensen et al. (2012) and Fredriksen and Stølen (2014), the calculations were restricted to horizontal distribution of old-age benefits from the new model for accrual of entitlements in 2050. By this simplification one can conveniently ignore indexation and actuarial adjustment through the flexible pension scheme. These elements of the reform have only small effects on the horizontal distribution of pension benefits. If distributional analyses are made after behavioural effects are included, these analyses will be a poor approximation for changes in welfare because they also reflect a voluntary shift in retirement ages Figure 8.

Figure 8

Download asset Open asset

A more favourable model for accrual of entitlements, before considering lower indexing of benefits than wage growth and adjustments for increasing life expectancy, means that no one seems to be worse off. While the figure gives a good indication of horizontal distributional consequences, it is misleading regarding the level of entitlements. Limited by the simplifications in these horizontal analyses the lower deciles will gain from the reform while accrual of benefits for old-age pensioners between the second and the fifth deciles is almost unaffected. The top five pension income deciles will experience an increase in entitlements, reflecting a somewhat increased accrual coefficient, and full accumulation of entitlements between 6 and 7.1 BPU with the new system, against only one-third with the old. The improvement for these deciles is larger for men than for women.

Christensen et al. (2012) also present overall distributional effects including tightening elements for the reform of old-age pensions in NIS by calculating total old-age pension benefits over the period as pensioner relative to total lifetime earnings, by income group and gender. An updated version for the 1963-cohort, which is the first cohort entirely encompassed by the new system, is presented in Figure 9. When the components for life expectancy adjustment and lower indexation of benefits in payment are included, everyone receives lower total benefits relative to total lifetime earnings. But because life-expectancy adjustments are more lenient for minimum pensions, persons in the first decile are less hurt by the reform than others. Although persons with high labour market incomes also seem to lose less from the reform than persons with medium incomes, NIS for old-age pensions is still highly redistributive over the lifecycle. Because of longer life expectancy for women and elements in the system for accumulation of entitlements favouring women, NIS also redistributes incomes from men to women. After the reform total lifetime pensions are only around 20 percent of total lifetime earnings for men from the third decile and above.

Figure 9

Download asset Open asset

Gender issues were a key feature in the debate to reform the Norwegian pension system prior to 2011. As a result, the reformed National Insurance System has several redistributive mechanisms that aim to compensate women directly for caring obligations or indirectly though redistribution from high to low lifetime earners: a gender neutral annuity divisor, a ceiling on annual earnings, generous child credits, a possibility for widows/widowers to inherit pension rights from a deceased spouse, a targeted guarantee pensions with higher benefit rates to single pensioners compared to married/cohabitating pensioners, and finally the tax system that is particularly progressive in its treatment of pensioners and pension income.

Halvorsen and Pedersen (2019) study in more detail the distributional components built into the reformed system aimed at achieving greater gender equality. The analysis deviates from previous studies using MOSART by analyzing the outcomes of a stepwise introduction of redistributive elements under stylized assumptions for a given cohort. Table 2 shows each of the redistributive elements’ quantitative effect on (individual) average pension income over retirement for men and women respectively.

As shown by Table 2, a hypothetical actuarially fair pension system would produce a gender gap of 43%, while the actual reformed system produces a gender gap of only 7% after taking account of income taxation. The gender-neutral annuity divisor gives the single most important contribution to close the gender gap, and the effect would have been even larger if benefits had been fully indexed with wages over the retirement phase. Among those components that are classified as generally redistributive (from rich to poor) the income ceiling appears to be the most important for closing the gender gap followed by the guarantee pension and the tax system. An important caveat here is that the two latter components would have had a stronger impact without the other redistributive components in place.

The concept of pension wealth has been extensively used by economists in theoretical and applied research since Feldstein (1974), but in recent decades, due to population ageing, it has attracted the attention from a wider scientific community and from policy makers. Pension wealth constitutes a significant proportion of total household wealth, but few countries have a comprehensive measure of pension wealth at the individual level.

However, National Accounts practices are evolving, and statistical offices are now required to report on the value of all pension entitlements in a supplementary table. The collected information must be comparable between countries whether pension schemes are funded or not. While pension obligations for entitlements under the third pillar and for funded schemes earlier had been included in standard National Accounts, the new system especially introduces compiling statistics on accrued-to-date liabilities in unfunded schemes. The production of statistics on accrued pension wealth in schemes for which no actuarial estimations have previously been estimated has spurred new use of microsimulation models.

Accrued-to-date liabilities are currently being calculated by using microsimulation models in France, Estonia and Belgium. However, none of these countries use the microsimulation model to make individual estimates of pension wealth. In a new project we use MOSART to calculate accrued to date liabilities -- or accrued pension wealth -- at the individual level. The goal is to use the data for household pension wealth statistics, and to analyze the distribution of individual pension wealth data and how it relates to the distribution of other personal wealth (both financial and real wealth).

Wealth data has been collected that covers all three pillars; National insurance, occupational pensions (both public and private) and personal pension saving. The actual value is used for members of defined contribution (DC) plans, but MOSART is used for calculating the discounted sum of benefits from defined benefit (DB) plans. Using a microsimulation model, pension wealth may be calculated under realistic and detailed assumptions about job market behavior, choices of retirement age and individual expected longevity (based on socio-demographic characteristics). As pointed out by Blanchet et al. (2017) the complexity of the pension system implies that an approach using dynamic micro-simulation seems to be far more efficient in projecting pension entitlements than an aggregate approximation. Micro-simulation allows a detailed account of the pension legislation at the individual level, applied to a heterogeneous population. An additional strength of the MOSART model is that the individual work histories are quite accurately embedded in the model due to long panels of register data. Furthermore, due to the Norwegian pension reform introduced from 2011, accumulation of entitlements from both the old and the new system can be considered.

Since pension wealth in defined benefit and non-funded defined contribution plans is the present value of expected pension streams, this involves making assumptions about discount rates and survival probabilities. To compute individual survival probabilities, most previous studies are generally based on official life tables. As described before, MOSART has a more sophisticated modelling of survival probabilities that also account for socioeconomic differences. Furthermore, the project involves using the model in a slightly different way. To get an unbiased estimator of longevity (and subsequently pension wealth) at the individual level we need to simulate many lifetimes for each individual.³ Given a full population this is computationally demanding. Therefore, we simulate the model 100 times using stratified, conditional and proportional sampling to increase efficiency. The result is as if we ran 2,000–10,000 ordinary simulations per person. Results from the project are reported in Smogeli and Halvorsen (2019) and Fredriksen and Halvorsen (2019).

MOSART is one of the longest running and actively used models in the family of dynamic microsimulation models. Its main use has been projections and policy analyses regarding the design and evaluation of the Norwegian pension system, but the model is also used for projecting labour force by educational level and for construction of pension wealth statistics. MOSART was important towards the end of the 1990s in illuminating the need for a reform and is still important in the evaluation and further adjustments of the pension system.

While similar models are used for analysing effects on pension expenditures from reforms of the system, a heavy weight in the Norwegian debate has been put on adequacy and distributional effects. This is also reflected in the design and the analyses made. By including relevant labour supply effects and incorporating these effects in a general equilibrium model it has been possible to show that positive effects on labour supply from the reform may be even more important for improving future fiscal sustainability than the tightening effects on pension expenditures.

Econometric analyses of labour supply show that there is a strong positive correlation between the level of education and performance in the labour market as well as the age of retirement. Modelling of educational behaviour is thus a main strength of MOSART compared with corresponding models in many other countries. Being situated in a statistical office with easy access to relevant micro data is also a great advantage, and simulations with the model are based on information from the entire Norwegian population. In addition, being situated in a Research Department with close connection to other research institutes as well as use of the results in important Ministries is of great value for securing quality and relevance.

Because of heavy use of the model in analysing the reform in combination with a limited staff, too few resources have been used in technical maintenance and updating of the transition probabilities. If re-estimation of the transition probabilities had been integrated into the running of the model, this would have made maintenance more efficient and significantly improved the quality of the parameters. If we have had the possibility to start over with today’s access to data and computer capacity, we would probably have chosen month as time-unit instead of year. With this choice, several of the existing weakness in aggregating over the year would have been eliminated

1. 1
   Building and interpreting Macro/Micro estimates of Accrued-to-Date pension liabilities: French reforms as a case study

   1. D Blanchet
   2. S Le Minez
   3. A Marino

   (2017)

   Review of Income and Wealth 63:70–94.

   https://doi.org/10.1111/roiw.12207

   • Google Scholar
2. 2
   Life expectancy and claiming behavior in a flexible pension system

   1. CN Brinch
   2. D Fredriksen
   3. OL Vestad

   (2018)

   The Scandinavian Journal of Economics 120:979–1010.

   https://doi.org/10.1111/sjoe.12271

   • Google Scholar
3. 3
   Cohort and Period Fertility for Norway 1820-1993. Reports 1994/27

   1. H Brunborg
   2. SE Mamelund

   (1994)

   Statistics Norway.

   • Google Scholar
4. 4
   Forecasting immigration in official population projections using an econometric model

   1. Ådne Cappelen
   2. T Skjerpen
   3. M Tønnessen

   (2015)

   The International migration review 49:945–980.

   • Google Scholar
5. 5
   The association between income and life expectancy in the United States, 2001-2014

   1. R Chetty
   2. M Stepner
   3. S Abraham
   4. S Lin
   5. B Scuderi
   6. N Turner
   7. A Bergeron
   8. D Cutler

   (2016)

   JAMA 315:1750–1766.

   https://doi.org/10.1001/jama.2016.4226

   • Google Scholar
6. 6
   Pension Reform in Norway: Combining NDC and Distributional Goals

   1. AM Christensen
   2. D Fredriksen
   3. OC Lien
   4. NM Stølen

   (2012)

   In: R Holzmann, E Palmer, D Robalino, editors. Nonfinancial Defined Contribution Pension Schemes in a Changing Pension World: Progress, Lessons, and Implementation. © Washington, DC: World Bank. pp. 129–154.

   • Google Scholar
7. 7
   What are the consequences of the AWG projections and hypotheses on pension adequacy?

   1. G Dekkers
   2. R Conti
   3. R Desmet
   4. O Sundberg
   5. K Van den Bosch

   (2018)

   Brussels: Federal Planning Bureau.

   • Google Scholar
8. 8
   Applications of microsimulation modelling, Budapest, Társadalombiztositási Könyvtár

   1. G Dekkers
   2. K Van den Bosch

   (2016)

   13–33, Prospective microsimulation of pensions in European Member States, Applications of microsimulation modelling, Budapest, Társadalombiztositási Könyvtár.

   • Google Scholar
9. 9
   Social security, induced retirement, and aggregate capital accumulation

   1. M Feldstein

   (1974)

   Journal of Political Economy 82:905–926.

   https://doi.org/10.1086/260246

   • Google Scholar
10. 10
    Projections of Population, Education, Labour Supply and Public Pension Benefits. Analyses with the Dynamic Microsimulation Model MOSART. Social and Economic Studies 101

    1. D Fredriksen

    (1998)

    Statistics Norway.

    • Google Scholar
11. 11
    The distribution of pension wealth in Norway

    1. D Fredriksen
    2. E Halvorsen

    (2019)

    The distribution of pension wealth in Norway, Working paper.

    • Google Scholar
12. 12
    Fiscal effects of the Norwegian pension reform – a micro-macro assessment

    1. D Fredriksen
    2. E Holmøy
    3. B Strøm
    4. NM Stølen

    (2017)

    Journal of Pension Economics and Finance pp. 1–36.

    • Google Scholar
13. 13
    93–108, Gender Aspects of the Norwegian Pension System, Ashgate, New Pathways in Microsimulation

    1. D Fredriksen
    2. NM Stølen

    (2014)

    93–108, Gender Aspects of the Norwegian Pension System, Ashgate, New Pathways in Microsimulation.

    • Google Scholar
14. 14
    An empirical investigation of labor income processes

    1. F Guvenen

    (2009)

    Review of Economic Dynamics 12:58–79.

    https://doi.org/10.1016/j.red.2008.06.004

    • Google Scholar
15. 15
    Closing the gender gap in pensions: a microsimulation analysis of the Norwegian NDC pension system

    1. E Halvorsen
    2. AW Pedersen

    (2019)

    Journal of European Social Policy 29:130–143.

    https://doi.org/10.1177/0958928717754296

    • Google Scholar
16. 16
    Chapter 7: Earnings equations and rates of return: The Mincer equation and beyond

    1. JJ Heckman
    2. A Layne-Farrar
    3. PE Todd

    (2006)

    In: EA Hanushek, F Welch, editors. Handbook of the Economics of Education. Amsterdam: Elsevir. pp. 307–458.

    • Google Scholar
17. 17
    Pension reform and labor supply

    1. E Hernæs
    2. S Markussen
    3. J Piggott
    4. K Røed

    (2016)

    Journal of Public Economics 142:39–55.

    https://doi.org/10.1016/j.jpubeco.2016.08.009

    • Google Scholar
18. 18
    Socioeconomic inequalities in mortality among elderly people in 11 European populations

    1. M Huisman
    2. AE Kunst
    3. O Andersen
    4. M Bopp
    5. J-K Borgan
    6. C Borrell
    7. G Costa
    8. P Deboosere
    9. G Desplanques
    10. A Donkin
    11. S Gadeyne
    12. C Minder
    13. E Regidor
    14. T Spadea
    15. T Valkonen
    16. JP Mackenbach

    (2004)

    Journal of Epidemiology & Community Health 58:468–475.

    https://doi.org/10.1136/jech.2003.010496

    • Google Scholar
19. 19
    Household Projections for Norway 1990-2020, Part 1: Macrosimulations. Reports 1995/21

    1. N Keilman
    2. H Brunborg

    (1995)

    Statistics Norway.

    • Google Scholar
20. 20
    Association of household income with life expectancy and cause-specific mortality in Norway, 2005-2015

    1. JM Kinge
    2. JH Modalsli
    3. S Øverland
    4. HK Gjessing
    5. MC Tollånes
    6. AK Knudsen
    7. V Skirbekk
    8. BH Strand
    9. SE Håberg
    10. SE Vollset

    (2019)

    JAMA 321:1916–1925.

    https://doi.org/10.1001/jama.2019.4329

    • Google Scholar
21. 21
    Simulating an Ageing Population: A Microsimulation Approach Applied to Sweden

    1. A Klevmarken

    (2008)

    31–54, Dynamic Microsimulation for Policy Analysis: Problems and Solutions, Simulating an Ageing Population: A Microsimulation Approach Applied to Sweden, Emerald: Series Contributions to Economic Analysis, Bingley.

    • Google Scholar
22. 22
    Befolkningsframskrivingene 2018. Modeller, forutsetninger OG resultater

    1. S Leknes
    2. S Løkken
    3. A Syse
    4. M Tønnessen

    (2018)

    Reports 2018/21, Statistics Norway.

    • Google Scholar
23. 23
    A survey of dynamic microsimulation models: uses, model structure and methodology

    1. J Li
    2. C O’Donoghue

    (2013)

    International Journal of Microsimulation 6:3–55.

    https://doi.org/10.34196/ijm.00082

    • Google Scholar
24. 24
    Handbook of Microsimulation Modelling

    1. J Li
    2. C O’Donoghue
    3. G Dekkers

    (2014)

    305–343, Dynamic Microsimulation, Handbook of Microsimulation Modelling, 293 edn.

    • Google Scholar
25. 25
    A Microsimulation Approach for Modelling the Future Human Capital of EU28 Member Countries. WP-17-017, IIASA

    1. G Marois
    2. P Sabourin
    3. A Bélanger

    (2017)

    A Microsimulation Approach for Modelling the Future Human Capital of EU28 Member Countries. WP-17-017, IIASA.

    • Google Scholar
26. 26
    Modernisert folketrygd

    1. NOU

    Bærekraftig pensjon for framtida. The Ministry of finance and Ministry of social Affairs, Oslo (in Norwegian).

    • Google Scholar
27. 27
    A new type of socio-economic system

    1. GH Orcutt

    (1957)

    The Review of Economics and Statistics 39:116–123.

    https://doi.org/10.2307/1928528

    • Google Scholar
28. 28
    Microanalytic Simulation Models to Support Social and Financial Policy

    1. GH Orcutt
    2. J Merz
    3. H Quinke

    (1986)

    New York: North Holland.

    • Google Scholar
29. 29
    Dynamic Microsimulation: a methodological survey

    1. C O’Donoghue

    (2001)

    Brazilian Electronic Journal of Economics 4:77.

    • Google Scholar
30. 30
    Handbook of Microsimulation Modelling

    1. C O’Donoghue

    (2014)

    Emerald Group Publishing Limited.

    • Google Scholar
31. 31
    Increasing the impact of dynamic Microsimulation modelling

    1. C O’Donoghue
    2. G Dekkers

    (2018)

    International Journal of Microsimulation 11:61–96.

    https://doi.org/10.34196/ijm.00174

    • Google Scholar
32. 32
    A Microsimulation Model for Educational Forecasting

    1. NE Rasmussen
    2. P Stephensen

    (2014)

    A Microsimulation Model for Educational Forecasting, DREAM Working Paper Series 201402.

    • Google Scholar
33. 33
    Pensjonsformue i Norge 2017, Reports 2019/28

    1. PO Smogeli
    2. E Halvorsen

    (2019)

    Statistics Norway.

    • Google Scholar
34. 34
    Befolkningens utdanningsnivå, etter spørreundersøkelsen om utdanning fullført i utlandet. Reports 2013/24

    1. A Steinkellner
    2. AMR Holseter

    (2013)

    Statistics Norway.

    • Google Scholar
35. 35
    International Standard Classification of Education ISCED 2011

    1. Unesco Institute for statistics

    (2012)

    Montreal: Unesco Institute for Statistics.

    • Google Scholar
36. 36
    Preface – Orcutt’s vision 50 years on

    1. M Wolfson

    (2009)

    In: A Zaidi, A Harding, P Williamson, editors. New Frontiers in Microsimulation Modelling. Farnham: Ashgate. pp. 21–30.

    • Google Scholar

Author details

1. Leif Andreassen

   Statistics Norway, Oslo, Norway

   Competing interests

   No competing interests reported

2. Dennis Fredriksen

   Statistics Norway, Oslo, Norway

   Competing interests

   No competing interests reported

3. Hege M Gjefsen

   Statistics Norway, Oslo, Norway

   Competing interests

   No competing interests reported

4. Elin Halvorsen

   Statistics Norway, Oslo, Norway

   Competing interests

   No competing interests reported

   "This ORCID iD identifies the author of this article:" 0000-0003-1752-9595

5. Nils M Stølen

   Statistics Norway, Oslo, Norway

   For correspondence

   Nils.Martin.Stolen@ssb.no

   Competing interests

   No competing interests reported