Enhancing microsimulation analysis of wealth-related policies in EUROMOD

The Eurosystem Household Finance and Consumption Survey (from now on HFCS) is conducted in a comparative way across the Euro Area by the national banks and some statistical institutes and coordinated at the European Central Bank. It covers detailed household wealth, gross income and consumption information and therefore provides more information on wealth than the current used database in EUROMOD, namely the “European Union Survey of Income and Living Conditions” (EU-SILC). The latter is the standard database used for the analysis of poverty and inequality in the European Union.

A significant advantage of the HFCS is that the wealthy population is oversampled (except in Italy, Ireland and Finland), i.e. households that are situated at the higher end of the income and/or wealth distribution are more accurately covered in the sample. As argued by Davies et al. (2011) this matters because those households are less likely to participate in surveys and more likely to underreport, in particular when it comes to financial assets. Another interesting feature of the HFCS data is that it uses a multiple imputation technique to deal with selective item non-response. Since EUROMOD requires that there is no missing information, this has the major advantage that missing information does not need to be imputed by researchers themselves. For more information on the oversampling and multiple imputation procedure of the HFCS we refer to HFCN (Eurosystem Household Finance and Consumption Network) (2016).

There are two main options for constructing a EUROMOD input database including wealth information. The first is to impute wealth information from the HFCS into the existing EU-SILC based EUROMOD dataset, while another option is to build a completely new input dataset fully based on HFCS. Since we want to maintain the strengths of the HFCS in terms of oversampling and multiple imputation and because the HFCS covers in general all other information needed in a EUROMOD input dataset we chose the latter option. Hence, this means that we create an input dataset based on HFCS covering all ‘standard’ variables needed for tax-benefit simulations also included in EU-SILC (with a few exceptions) complemented with additional wealth-related variables. In order to utilise the multiple imputation advantage to the fullest, we created five separate input databases, each representing information on one of the five imputations. Each dataset is run through EUROMOD, the results shown in this paper represent the average over the five output datasets. Figari et al., 2007 noted that a database must fulfil certain requirements in order to be used in a sensible way in EUROMOD. As discussed by Kuypers et al. (2016), the HFCS fulfils the majority of these criteria such that it can be used in a reasonable way as input data in EUROMOD.

We constructed a EUROMOD input database based on the HFCS for a selection of 17 out of the 20 countries who participated in the second wave.¹ The Netherlands, Malta and Latvia are not included, mainly because of their small sample sizes. Moreover, given that the Netherlands has administrative wealth data which show a large discrepancy with the HFCS estimates (see Salverda, 2015), it does not seem appropriate to include it. Nevertheless, even without these countries our selection represents a broad range of European countries with various kinds of tax-benefit systems in place. Besides the fact that they differ in the more ‘traditional’ tax-benefit instruments (i.e. personal income tax, social insurance contributions, social transfers), they also largely differ in the extent and progressivity of wealth-related taxation or the way in which wealth is taken into account in determining eligibility for social transfers (i.e. ‘asset-testing’). The broad coverage of countries takes advantage of the harmonised framework of EUROMOD and allows to study the budgetary and redistributive effects of wealth-related taxation and other policies from a cross-country perspective.

Table 1 provides an overview of the income reference years of the input datasets for the respective countries. For most countries it is 2013 or one year before or after. For Spain it is 2010.²

In Table 2 we present sample characteristics of the EUROMOD input dataset based on HFCS (called ‘EM-HFCS’) in comparison with those of the dataset based on EU-SILC (called ‘EM-SILC’) which is closest to the HFCS income reference period. The sample size of EM-HFCS ranges from 1,289 households (4,223 individuals) in Cyprus to 12,035 households (28,845 individuals) in France. In most countries the sample size of EM-HFCS is (much) smaller than that of EM-SILC, which is also reflected in the higher value of average weights. Exceptions are France and Ireland, where EM-HFCS has a larger sample. In Finland the same set of people are covered in EU-SILC and HFCS. Following common EUROMOD conventions, children that were born after the end of the income reference period are removed from the sample in the input database. We only know the age of the individuals at the time of the interview and not the year in which they were born. Hence, we assume all individuals younger than one year old to be born after the income reference period. The outcomes of applying this procedure to the HFCS input data for each of the separate countries is presented in the column ‘Restricted individuals’.

The EUROMOD input dataset that is constructed based on the HFCS contains both variables which are also included in the EU-SILC database in order to simulate the ‘standard’ EUROMOD policies, as well as new input variables in order to simulate wealth-related taxes and policies. The required variables depend on the design of the specific taxes and policies. In the preparation of these input data several challenges needed to be addressed, mainly because of a lack of information in the original HFCS data. We focus here on issues which are applicable to most (or all) countries, for detailed country-specific issues we refer to Boone et al. (2019).

Survey data are generally available to researchers only after a considerable time lag. In case of the HFCS, data are usually available three years after the interviews take place. Hence, we would like to use the input data both for simulations of the policies as they existed in the income reference year as well as for more recent years. At the moment the most recent coding of the wealth-related taxes and policies in EUROMOD applies to the situation in 2017. To be able to run the input data from the second HFCS wave on the 2017 policies we need to uprate monetary variables to the price levels of 2017. Income components and other variables are uprated using the standard uprating indices included in EUROMOD (for more information on the general uprating procedure see the EUROMOD Country Reports). For monetary variables that are new to the EUROMOD input database we have constructed new uprate indices. As for EU-SILC non-monetary variables are assumed to have stayed the same.

We illustrate in Table 4 the construction of the new uprating indices with the example of Germany for which variables need to be uprated from 2013 to 2017. Detailed information on the uprating procedure for each country can be found in Boone et al. (2019). First, the main asset variables are uprated based on their respective aggregates in the national accounts⁵. In the case of Germany these were taken from the Federal Statistical Office Germany (2018a); Federal Statistical Office Germany (2018b) and the Deutsche Bundesbank (2018). Although categories of the national accounts and HFCS do not always coincide perfectly (Kavonius and Honkkila, 2013; Waltl, 2020), they are the best available information to take into account the evolution of assets and debt. We try to match the categories as close as possible. For self-employment business assets we used the categories ‘machinery & equipment’ and ‘intellectual property rights’ from the national accounts as proxy. For the HFCS asset categories ‘managed accounts’ and ‘money owed to households’ there is no information available in the national accounts. For managed accounts we apply the same uprating index as for mutual funds and for money owed to the household we use the EUROMOD default, i.e. the price index. The aggregate wealth variables ‘total financial assets’, ‘total real assets’ and ‘total assets’ are uprated by setting them equal to the sum of their uprated components. The uprate index for the value of real property is also applied to the value of real property at the time of purchase, which is used in the simulation of the real property transfer tax. Cadastral values are not uprated. Furthermore, the monetary variables used in the simulation of the inheritance and gift tax are uprated using administrative information on the total amount of inheritances and gifts larger than 0 euro (Federal Statistical Office Germany, 2018c). This information is not available for most countries, so then the uprate index is defined in terms of the evolution in government revenues from inheritance and gift taxation. Since the applicable tax legislation has not significantly changed in the period we are uprating over, it is relatively certain that changes in the tax revenues mainly reflect changes in the amount of wealth that is received or inherited.

As mentioned before we assume non-monetary information to have stayed the same. Yet, we simulate the real property transfer tax and the inheritance & gift tax only for those individuals who have experienced these events in the policy year, i.e. for the first only individuals who have bought a property in the policy year and for the second only indivuals who have received an inheritance or gift in the policy year. In the policy years after the income reference period we keep using the observations from the income reference period. Hence, in practice this means we replace the variable which contains information on the year a property is bought or an inheritance/gift is received with 2017 for those for which it is equal to the income reference period (i.e. 2013 is replaced by 2017 in the case of Germany).

Ownership of real property is taxed recurrently in all HFCS countries included in the analysis. The tax base differs between the different countries but can be divided into three separate categories. Most countries use the cadastral value of the property as tax base for the calculation of the property tax (Austria, Belgium, Finland, France, Germany, Italy, Luxembourg, Portugal and Spain). Other countries use the market value (Cyprus, Hungary, Ireland and Slovenia) or the property size in m² (e.g. Greece⁶, Hungary⁷, Poland and Slovakia) as tax base. In some countries there exist exemptions from the real property tax. Note that the Hungarian real property tax cannot be simulated since it requires detailed information at the municipality level that is not available in HFCS.

Transfers of real property are subject to a transfer tax that is payable by the buyer of the property in all countries. The purchase of immovable property is often preceded by taking out a mortgage, which is in some countries also taxed (Belgium, Italy, Portugal, Spain). With the exception of Italy, all countries levy the transfer tax on the price of the property (i.e. its fair market value), while in Italy the cadastral values are used as tax base. In general, there are no exemptions from this tax, although transfers of properties between lineal heirs or properties held by the government are exempt from taxation in some countries (e.g. Germany, Portugal, Spain …). For Finland and Slovakia, we are not able to simulate the transfer tax. For Finland this is because in HFCS information on wealth transfers is missing as it is based on a combination of register data and a supplementary module added to the EU-SILC survey (HFCN (Eurosystem Household Finance and Consumption Network), 2016). In Slovakia the real estate transfer tax was abolished in 2005, but there is still a very small registration fee in place. The simulation of this fee, however, requires information not available in HFCS. Because it consists of only a small fee, the budgetary importance of this is very limited.

Apart from Austria, Estonia and Slovakia inheritances and gifts are subject to taxation in all countries and are due by the beneficiary of the inheritance/gift. Overall, the value of the inheritance/gift is used as tax base. Often tax rates vary according to the kinship between the beneficiary and the deceased/donor with more favourable tax treatment for partners, descendants and ascendants compared to other relatives or non-related people. Inheritances and gifts are generally taxed in a progressive way, either through a progressive tax schedule (all countries except Hungary, Ireland, Italy, Luxembourg and Portugal) and/or by granting large allowances of several thousands of euros (Germany, Greece, Ireland, Italy). The inheritance and gift tax of Finland cannot be simulated which is again due to the missing information on wealth transfers mentioned above.

In the years for which the simulations are carried out, a general net wealth tax only existed in France⁸ and Spain⁹. In both countries the tax is levied on the net wealth (i.e. real and financial assets minus liabilities). It is levied on individuals who own a ‘high share of net wealth’, i.e. at least €1,300,000 in France and €700,000 in Spain (in the latter doubled for couples). Apart from the tax-free threshold, both countries have additional exemptions from the wealth tax included in their tax legislation. For instance, the value of the main residence is partially exempted and works of arts, antiques and retirement savings are fully exempted as well as business assets under certain conditions. Tax rates are progressive in both France and Spain. Italy levies a ‘specific net wealth tax’, which entails the taxation of bank accounts and financial assets.

In the simulation of all the wealth-related taxes we simulate the rules as they apply to residents and the wealth held in the country of residence. Other rules may apply to wealth held in the country by non-residents or the wealth held by residents in other countries.

Table 6 presents an overview of the policies which were already simulated in EUROMOD, but which have often been refined by taking into account (more detailed) wealth information where necessary. Again, we focus on the 2017 policies, but the situation is largely the same for the income reference period. We classify policies as ‘exists and simulated’ (ES), ‘exists, simulated and refined’ (ESR), ‘exists and not simulated’ (EN) and ‘does not exists’ (N). Taxation of income from financial assets and from real property is for all countries included in EUROMOD. Tax reliefs for mortgage repayments respectively for contributions to private pension funds are also well covered in the refined policies (except for Cyprus, Finland, and Slovakia, resp. for Germany). The same applies for asset tests for social benefits, where the asset test has either been added to the existing policy or the asset test that was coded was refined with additional information (except for Estonia and Slovenia). Country specific taxes are refined for Belgium and Cyprus. For Belgium this entails the ‘tax on long-term savings’ that is levied once people turn 60 years old, while for Cyprus the ‘special contribution to defense’ which is levied on income from financial assets is simulated.

Table 7 presents summary statistics of original & pension income and disposable income for EM-HFCS and EM-SILC, although we do not claim that one data is better than the other. We show here the results for the income reference year as these directly reflect the underlying databases, but the differences between EM-HFCS and EM-SILC are the same for 2017 as the same uprating indices are applied to the income variables for both datasets. To be able to compare accurately between EM-HFCS and EM-SILC the newly simulated wealth-related taxes are here not subtracted from disposable income. Note that the figures presented below refer to the mean over the five imputations. All figures are calculated based on the annual household disposable income, equivalised by the OECD modified scale and all individuals are included in the calculations.

We highlight the extent of the differences as follows: white cells refer to differences of less than 5%, light grey cells refer to differences of between 5% and 10%, medium grey cells refer to differences of between 10% and 20% and dark grey cells refer to differences of more than 20%. Comparability between the results of EM-HFCS and EM-SILC varies widely across countries. Results are close to each other for Finland, Portugal and Slovakia, while they diverge rather strongly for Austria, Estonia, France and Slovenia. Differences are usually larger for original & pension income than for disposable income and larger for the mean than for the median. The differences mostly reflect higher amounts in EM-HFCS than in EM-SILC, which might be related to the oversampling that is applied in the HFCS (see above). Gini coefficients are often also higher in EM-HFCS than in EM-SILC.

Since the differences between EM-HFCS and EM-SILC are sometimes relatively large and they vary widely across countries, this may potentially lead to different results of simulations of the impact of tax-benefit policies and the country rankings thereof. Nevertheless, the comparison only reveals differences, but not which dataset is closest to reality. Both datasets are household surveys having each important strengths but also suffering from weaknesses. Hence, at this moment there is no way of knowing which dataset provides the ‘best’ estimates of inequality, poverty and redistribution. Therefore, we argue to consider the datasets as complements rather than substitutes.

We now turn to the macro-validation of the newly added wealth policies in EUROMOD.

Table 8 summarises for each of the taxes the number of observations in the data which are in theory liable to pay the tax and the final observed number of actual taxpayers. The number of potentially liable observations refers to those observations in the input data that could be theoretically taxed. This number refers to all observations that possess the type of asset or have experienced the type of wealth transfer which is taxed in the respective tax, without taking into account any tax legislation or data constraints. This number does not necessarily correspond to the final number of actual taxpayers since units may not pay a tax for several reasons (e.g. missing input data, exemptions foreseen in tax legislation…). The criteria used to calculate the number of potentially liable observations is the same across all countries:

• Real property tax: households are considered potentially liable if they (partially) own at least one property.

• Real property transfer tax: households are considered potentially liable if they bought at least one property in the policy year.

• Inheritance tax: households are considered potentially liable if they received a positive inheritance in the policy year.

• Gift tax: individuals are considered potentially liable if they received a positive gift in the policy year.

• Mortgage registration duties: households are considered potentially liable if they took out at least one mortgage in the policy year.

• Net wealth tax: individuals are considered potentially liable if they own positive net wealth.

The final number of observed taxpayers are those cases that eventually pay a positive tax after taking into account the tax rules (weighted population numbers are presented between parentheses).

The number of observations differ strongly between the different wealth-related taxes and countries. In general, the number of potentially liable observations and the number of observed taxpayers is highest for the real property tax. In comparison, the number of observations for the inheritance & gift tax, real property transfer tax and mortgage registration duties are considerably lower. In general, the number of observed taxpayers is the same in 2017 compared to the income reference period as most tax legislation did not change (drastically) over this period. The most important differences are (1) that for Greece the number of payers of the real property tax increased between the income reference year and 2017 because of a reform from an ‘emergency property tax’ to the property tax as it exists today, (2) that for Italy the number of payers of the real property tax decreased because the main residence has become exempted from taxation since 2016 and (3) that the number of tax payers of the Spanish net wealth tax is zero in the income reference year (2010) since this tax was abolished between 2008 and 2011.

Table 9 presents the comparison of tax revenues for the simulated wealth-related taxes with external figures, mainly derived from the OECD Tax Revenue Database. We show here both the validation for the income reference period as for 2017 because the first shows best how well our simulations perform, while the latter provides some insight in the appropriateness of the uprating. Since the number of observations is highest for the real property tax this tax is on average the most accurately simulated wealth tax. For the other taxes the simulated revenues diverge more from the external figures. There are several reasons for this. First, as Table 8 showed the number of observations in the underlying data are often very low such that the results largely depend on a few cases. When these cases happen to be outliers this has a massive effect on the results. Second, as the results largely depend on the input database, underreporting in the HFCS data may have an effect on the simulation results. Moreover, our simulations are not always completely comparable to the external statistics as the latter are not always available at a detailed level. For instance, we simulate wealth taxes for households, but external figures often do not make a distinction between taxes paid by households versus other economic actors.

The HFCS is by itself not a suitable dataset for (re)distributive analyses as it only covers incomes gross of liabilities for taxes and social insurance contributions. Also, apart from pensions and unemployment benefits, all other social transfers are covered under a single variable. Including the HFCS as an underlying database for EUROMOD allows to simulate these components of the tax-benefit systems for the households covered in the HFCS and to derive disposable incomes. In this way the HFCS becomes an additional source which can be used for research on poverty, inequality and redistribution in Europe, which is currently largely based on EU-SILC and the Luxembourg Income Study (LIS). Moreover, compared to these sources the HFCS has the major advantage that it covers information on both income and wealth such that poverty, inequality and redistribution can be studied both in terms of the distribution of income as well as the distribution of wealth, or a combination of the two.

A first interesting application can be situated in the so-called ‘asset-based poverty’ literature. In this literature it is argued that financial well-being and precariousness depend on both income and wealth and hence that wealth should be taken into account when determining who is worse off (Kuypers and Marx, 2019). Given that the two distributions are not perfectly correlated poverty measures based on income alone tend to overstate poverty rates among households with low income, but median to high net wealth, while they potentially ignore the precarious situation of households with incomes above the poverty threshold, but with very low assets or bearing a large debt burden (Kuypers and Marx, 2021). Therefore, two approaches have been proposed to define joint income-wealth poverty indicators. Up until a few years ago this literature was largely US-oriented. The combination of disposable incomes and net wealth in the HFCS now also allows to estimate such indicators for Europe (see for instance Kuypers and Marx, 2021 forthcoming; Kuypers and Marx (2018)).

Another possible application is to study the redistributive effects of taxes and benefits. Table 10, for instance, shows the absolute and relative redistributive effect achieved by income taxes and social insurance contributions (SIC). The absolute redistributive effect is given by the difference between the Gini coefficient of the income distribution before and after income taxes and social insurance contributions are taken into account. The relative redistributive effect than expresses this absolute redistributive effect as a percentage of the Gini coefficient before income taxes and social insurance contributions are taken into account. A positive redistributive effect means inequality is reduced through taxes and SIC, while a negative sign indicates an increase in inequality. We find that income taxes and social insurance contributions reduce inequality by between 1.5% in Poland and 21.6% in Austria. The last two columns in Table 10 show the two building blocks of the redistributive effect of taxes and SIC, namely their size (average tax rate) and their progressivity (Kakwani index). The average tax rates vary between 10% in Spain and 35% in Belgium. The Kakwani index shows that the total of income taxes and social insurance contributions is progressive, most strongly so in Ireland and less so in Poland and Hungary.

In second instance, the EUROMOD-HFCS tool also provides the possibility to study new research questions. First, due to broader policy scope we cannot only study the redistributive effect of the ‘traditional’ tax-benefit instruments, but also see how redistributive wealth taxation currently is. It is possible to evaluate the redistributive effects of wealth taxation against the income distribution (see Table 10), against the wealth distribution or from a joint income-wealth perspective (for the latter two see Kuypers et al., 2020). Irrespective of the framework chosen, we find that in their current form wealth-related taxes are hardly redistributive. While Table 10 shows that wealth taxes are often regressive when assessed against the distribution of income (i.e. the Kakwani index is negative), they are more progressive when assessed against the (joint) distribution of wealth (and income). However, even when the latter perspectives are taken wealth-related taxes do not achieve any significant redistribution as a consequence of their very small size (Kuypers et al., 2020). The analysis of the wealth taxes can be combined with the standard redistributive analyses as described above. The whole tax-benefit system can then be evaluated against the joint distribution of income and wealth as is done in Kuypers et al. (2019); Kuypers et al. (2021). This exercise shows that European welfare states are not as redistributive as they are generally believed to be as most efforts go towards reducing income inequalities, while wealth inequalities remain largely unaddressed. Besides studying current wealth tax systems, it is also possible to simulate the budgetary and (re)distributive effects of recent wealth tax proposals, or explore the prospects of radical new ideas (for instance taxing a wealth annuity in the personal income tax, see Kuypers et al., 2020). Also, it allows to study these reforms in accordance with other tax-benefit instruments.

Another new research question is related to how wealth is taken into account in determining the eligibility conditions for means-tested benefits. Many European countries have such asset tests in place. Analysing the asset test in minimum income protection schemes in EU member states Marchal et al. (2021) distinguish between two main types of asset tests. The first and most prevalent type applies a disqualification threshold, i.e. when assets are above a certain threshold applicants become immediately ineligible. The second type takes assets into account in a given percentage or at a fictional rate of return which is added to the income in the means-test, such that applicants are eligible to lower minimum income benefits as more assets are available (and eventually at high asset levels also become fully ineligible). Yet, this rate of return is usually higher than actual returns received so in practice assets often also need to be realised. The combination of HFCS with EUROMOD allows to simulate benefit eligibility with and without taking into account the asset test. Hence, it shows the effect of asset testing on eligibility rates, poverty rates and budgets and whether these are different for the two types of asset tests (Marchal et al., 2021).

A final example relates to public encouragement of wealth accumulation. Many countries across Europe have put in place tax expenditures for instance for mortgage interests, private pension savings and financial income. Using the HFCS data and EUROMOD we can analyse who benefits from these tax expenditures and its cost in terms of forgone tax revenues. Results indicate that these tax expenditures are regressive instruments. Poor households hardly benefit because they do not have the means to invest in the types of assets that are encouraged, because they do not pay sufficient taxes to be able to benefit from a deduction/credit and/or because they are discouraged to save because of asset testing. The EUROMOD-HFCS tool then also allows to simulate new proposed policies, for instance, subsidising wealth accumulation among the poor (see Kuypers, 2018).

In the first wave 15 countries participated, the recently released third wave took place in 22 countries.

In the new release of the HFCS data in March 2020, the Spanish wave with income reference period 2010 was moved to the first wave and replaced by results with income reference period 2014. Our databases were constructed on a previous version of the second wave data.

The validation only provides information on the consistency of the average level of the cadastral values, and hence the total tax base, but not their distribution. By applying the same ratio to all properties we might suppress the variance in the variable used for the simulations compared to reality. Yet, the relationship between market and cadastral values is not straightforward due to the lack of revisions of cadastral values in most countries. Therefore, there is no way of knowing a priori to which extent such distributional bias is actually induced.

It is possible that households change their asset portfolio by swapping between different types of assets, such that we may in some cases subtract amounts which were part of net wealth on the first of January in the income reference period. We, however, assume that for the majority of the households buying a house represents a new type of wealth.

Since the wealth of those at the top of distribution often increases at a faster pace than at the bottom, uprating the wealth of all households by the same index might reduce the inequality included in the simulations as compared to reality. Yet, by applying different uprating indices to the separate wealth components part of this differential increase is captured as those at the top of the wealth distribution more often invest in financial assets such as listed shares, while those at the bottom own most of their wealth in deposits and value of their main residence.

Property size is combined with information on different coefficients (building age coefficient, floor or house coefficient, façade coefficient and incomplete building coefficient) to determine the tax base.

In Hungary, either the property size or adjusted market value can be used as tax base, depending on the municipality.

Replaced by a tax levied only on real estate wealth since 1^st of January 2018.

Abolished between 2008 and 2011, but thereafter reintroduced.

This research was carried out as part of the projects “Update and Extension of the EUROMOD Wealth Taxation project” (JRC/SVQ/2017/B.2/0018/NC) funded by the Joint Research Centre – Fiscal Policy Analysis Unit (Sevilla) of the European Commission and the “European Union’s Horizon 2020 research and innovation programme” under grant agreement No.730998 (InGRID-2).

We gratefully acknowledge the support, guidance and comments received by the staff of the Fiscal Policy Analysis Unit, in particular Andreas Thiemann. The analysis builds on data from the second wave of the Eurosystem Household Finance and Consumption Survey (HFCS) and the EU-wide micro-simulation model EUROMOD (version H1.0). Members of the HFCN and the EUROMOD framework are gratefully acknowledged for their contributions. Results may not correspond to those of the data providers.

1. Version of Record published: December 31, 2020 (version 1)

© 2020, Kuypers, Boone, Derboven, FigariVerbist

This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.