The country factor on regional income distributions in Europe: A functional ANOVa approach

The country factor on regional income distributions in Europe: A functional ANOVA approach Pedro Delicado1 and Magda Mercader2 (First version: July 2005. This version: June 2006) Abstract The distribution of regional Gini indices in Europe using the income distribution before taxes and transfers is not explained by the country to which the region belongs, i.e. the dispersion of the Ginis is not significantly reduced when we control for the country variable. On the contrary, there is a clear dependency between the regional Ginis and the country when the distribution of income before taxes and transfers is considered. This evidence is based on EUROMOD a multicountry tax-benefit model of the EU-15 (See Mercader and Levy 2004). We study to what extent this conclusion holds when we consider the complete income distributions instead of a summary inequality measure such as the Gini index. We use functional ANOVA (following Cuevas et al. 2004) in order to study the country explicative power on the dispersion of regional income density functions (estimated non-parametrically) before and after taxes and transfers. Our statistical evidence suggests that regional income distributions in different countries are different, both before and after redistribution takes place. However, the null assumption of equality of mean regional distributions among countries (a factor country equal to zero) is rejected more strongly in the after distribution case. Acknowledgements In this paper we use data from the EUROMOD project. We are indebted to all past and current members of the EUROMOD consortium. Our acknowledgment is also for Horacio Levy who helped us with the data. The views expressed in this paper, as well as any errors, are the responsibility of the authors.3 1 Universitat Politècnica de Catalunya. Pedro.Delicad[email protected] 2 Universitat Autònoma de Barcelona. Spain. [email protected] 3 In particular, this applies to the interpretation of EUROMOD results and any errors in its use. EUROMOD is continually being improved and updated and the results presented here represent work in progress. The EUROMOD version used in this work is 25a. EUROMOD relies on micro-data from 15 different sources for fifteen countries. These are: the European Community Household Panel (ECHP) User Data Base made available by Eurostat, the Austrian version of the ECHP facilitated by the Interdisciplinary Centre for Comparative Research in the Social Sciences, the Panel Survey on Belgian Households (PSBH) contributed by the University of Liège and the University of Antwerp, the Income Distribution Survey given by Statistics Finland, the Enquête sur les Budgets Familiaux (EBF) made available by INSEE, the public use version of the German Socio Economic Panel Study (GSOEP) supplied by the German Institute for Economic Research (DIW), Berlin, the Living in Ireland Survey contributed by the Economic and Social Research Institute, the Survey of Household Income and Wealth (SHIW95) given by the Bank of Italy, the Socio-Economic Panel for Luxembourg (PSELL-2) facilitated by CEPS/INSTEAD, the Socio-Economic Panel Survey (SEP) made available by Statistics Netherlands through the mediation of the Netherlands Organisation for Scientific Research - Scientific Statistical Agency, the Income Distribution Survey supplied by Statistics Sweden and the Family Expenditure Survey (FES), offered by the UK Office for National Statistics (ONS) through the Data Archive. Material from the FES is Crown Copyright and is used by permission. Neither the ONS nor the Data Archive bear any responsibility for the analysis or interpretation of the data reported here. An equivalent disclaimer applies for all other data sources and their respective providers cited in this acknowledgement. 2 1. Introduction Countries, through different mechanisms, including their history and political, social and economic institutions, are important factors in distinguishing final personal inequality levels. One of the key elements in understanding of the shape of a country’s final income distribution is its tax-benefit system (See Atkinson, 2000). Cross-country empirical evidence shows how the impact of taxes and transfers critically depends on the country (See Atkinson et al (1995), Wagstaff et al (1999) among others). This paper aims at providing further evidence on the significance of national tax-benefit systems in explaining the final shape of income distributions. Relying on microdata from EUROMOD, representative of the European income distribution before and after the operation of tax-benefit systems4, an experimental statistical approach is adopted. We analyse the role of the country factor in explaining the variability in regional income distributions shapes before and after national tax-benefit systems operate. The main questions we aim at answering are: Which is the contribution of the country factor on the observed variability of regional income distributions before taxes and transfers? By how much this contribution increases when the role of taxes and transfers systems has been accounted for? In a recent paper Mercader and Levy (2004) study similar questions. By means of a simple ANOVA model they show that while the distribution of regional Gini indices in Europe using the income distribution before taxes and transfers is not significantly explained by the country to which the region belongs, i.e. the dispersion of the Ginis is not significantly reduced when we control for the country variable, there is a clear dependency between the regional Ginis and the country when the distribution after taxes and transfers is considered. In this paper we take further their analysis. We study to what extent this conclusion holds when we consider the complete income distributions instead of a summary inequality measure such as the Gini index. In order to take account of the full income distribution, we use Functional Data Analysis techniques.5 Functional ANOVA (following Cuevas et al. 2004) is used in order to study the country explicative power on the dispersion of regional income density functions (estimated non-parametrically) before and after taxes and transfers. If in micro-simulation it is common the evaluation of the effect of a marginal policy change on a given population, this work uses a complementary approach. We aim at exploring what can be learned from applying the same system on marginally different populations. Regions in a given country offer a particularly attractive natural framework for experimental analysis. Firstly, because regions can be seen as being distorted pictures of a given country. Regions belonging to a country often share a number of characteristics such as a common legal reference system, similar labour market and demographic structures and traditions. Secondly, regions are real sub-populations of a given country. Knowing the effect of a country tax-benefit system in one of its regions is therefore of clear relevance, also from a normative perspective. One could, of course, argue that in a European context with 25 countries, having each its own tax-benefit 4 Only countries from the European Union before May 2004 are considered. 5 Jenkins and Van Kerm (2004) using similar techniques show the relevance of looking at the whole distribution in accounting for income distribution trends. 3 system, the ideal experimental structure for our analysis would require a matrix in which we evaluate the 25 tax-benefit-systems on the 25 countries. However, we do not have this information available. Microsimulation research has demonstrate that applying to country k the tax-benefit system of country j poses a number of difficult issues that cannot be always easily solved (Atkinson et al, 1988). Our experiments, by relying on real populations, avoid adopting arbitrary assumptions that such an ideal experimental structure would require. The organisation of the paper is as follows. After this introduction, Section 2 includes a description of EUROMOD, the tax-benefit model used, presents the regional data as well as other assumptions underlying the construction of the income distributions studied in this work. Section 3 summarises previous findings based on the univariate case for regional Ginis before and after the role of national tax-benefit systems. Description of both the Functional Data Analysis techniques for dealing with income density functions and the functional ANOVA model for estimating the country effect on regional income distributions is offered in Section 4. This section also explains the methods proposed for testing the null hypothesis of equality of average Ginis among countries. The procedure used for the estimation of regional density functions is offered in Section 5. Results of the functional ANOVA on regional EU density functions are presented in Section 6. We end with a concluding section with some policy implications from our analysis. 2. A description of the data This analysis relies on EUROMOD. EUROMOD is an integrated tax-benefit micro simulation model for all countries of the old European Union before May 2004.6 EUROMOD is a source of harmonized micro-data on the different income components. It allows to construct in a comparable manner the income distribution before and after the main taxes and cash benefit programs in place in the different countries. The original databases on which EUROMOD relies are summarised in Table 1. <Insert Table 1> Our analysis focuses on the comparison of two income distributions. The distribution of income before public cash transfers are added and taxes are deducted. More precisely, the before taxes and transfers income distribution includes all components of market income: wages and salaries and self-employment income (net of employer insurance contributions and other benefits, but gross of employee contributions to such schemes), property income (interest, rents, dividends) as well as occupational pensions from employers, regular inter-household cash transfers and other sources of income which are not government transfers. The after taxes and transfers income distribution is the before income distribution plus all social transfers in cash minus employee Social Insurance Contributions, personal income taxes and other taxes. Indirect taxes are not considered. All income distribution estimates presented are based on the relative equivalent income (net or after taxes and transfers) per person in each region. Incomes are equivalent incomes in the sense that the household incomes are divided by the equivalent number of adults living in there (using the modified OECD scale) and income per person 6 For details about the EUROMOD’s team and project, see its website: http://www.econ.cam.ac.uk/dae/mu/emod. For a detailed description of the model see Sutherland (2001). 4 indicates that each household income is weighted by their number of members. Incomes are relative because in each region the observed equivalent incomes are divided by the median of regional equivalent incomes. The analysis does not look at the variability of income distributions due to differences in income levels but it focuses on the differences in shapes. Finally, the whole analysis refers to 1998 annual incomes.7 Regional information in EUROMOD Table 2 shows the regional information used in our analysis. For most countries, the official classification of regions established by Eurostat (NUTS system) is available. Our analysis uses NUTS1 information for most countries but we keep NUTS2 for countries in which this information is available.8 Notice that average population size per region significantly diverges among countries, from less than 1 million people in Finland to 5.5 million people in Spain. Also, the number of regions varies greatly. Out of the total number of regions considered, France concentrates 22 of the regions analysed, while Austria and Belgium only 3 regions each and Greece 4. For the other countries, the number of regions is in the range 6 to 12. <Insert Table 2> Table 3 shows the sample size available for the different regions. Sample sizes vary also greatly, from 3,404 in Östra Mellansverige in Sweden to Corse in France for which only 38 observations are available. <Insert Table 3> 3. Previous findings Mercader and Levy (2004) study the role of the country in explaining the variability of regional Ginis in Europe before and after the application of national taxes and transfers systems. They use the same data we use here. One of their main finding is that distribution of regional Ginis indices in Europe using the before taxes and transfers income distribution is not explained by the country to which the region belongs, i.e. the dispersion of the Ginis is not significantly reduced when we control for the country variable. On the contrary, there is a clear dependency between the regional Ginis distribution and the country when the income distribution after taxes and transfers is considered. This is illustrated in Figures 1 and 2 showing a box-plot of regional Gini indices by country before and after redistribution takes place.9 7 The Danish, Swedish and UK currencies were converted into Euro using the exchange rate of December 31st, 1998. 8 NUTS2 is a more disaggregated classification than NUTS1. 9 The internal box line represents the median regional Gini and while the box upper and lower bands indicate the regional gini Gini at the 25th and 75th percentiles. The lines extending above and below the box report the minimum and the maximum ginis Ginis in each country and the symbol ° the outliers. 5 Figure 1. .4 .45 .5 .55 .6 Market Gini pt gr es it uk fr fi se at de be Source: EUROMOD Box Plot: Original Gini in EU regions, by country Note: Market Gini is calculated using the household income before taxes and transfers. The distribution is equivalised using the modified OECD scale. Each household is weighted according to the number of members. Taken from Mercader and Levy (2004). . T Figure 2. .2 .25 .3 .35 .4 Disposable Gini pt gr es it uk fr fi se at de be Box Plot: Disposable Gini in EU regions, by country Note: Disposable Gini is calculated using the household income after taxes and transfers. The distribution is equivalised using the modified OECD scale. Each household is weighted according to the number of members. Taken from Mercader and Levy (2004). 6 It is interesting to notice that while regional gross income inequality levels appear to be rather independent from the country in which regions belong, the level of disposable income inequality of a given region is to be observed as much more dependent on the country in which this region belongs. Most regions in Southern Europe (Portugal, Italy, Greece and Spain) show a Gini coefficient above .3. The UK regions follow with Gini coefficients around 0.3. Most French and Belgian regions are in the interval .25-.3. The lowest Gini are found in the Finnish regions and regions from Austria (values around .2- .25). There are significant differences in disposable inequality levels in regions in Sweden (between .25 and .35) and Germany (from around .2 to .3). The results of fitting a simple ANOVA model to the regional Ginis before and after the action of the tax-benefit allow to confirm these findings. The null hypothesis that average regional Ginis among countries are equal cannot be rejected on the Ginis before redistribution. In fact the country factor explains only 16% of the variance observed. On the other hand, the null hypothesis that the average Ginis among countries are the same is rejected on the after taxes and transfers case. In this second case, the country factor explains 69% of the variance of the Ginis. Thus, the country is, through the national taxbenefit system, a decisive factor on regions’ final income inequality levels while this is not so in the case of market income inequality (See Mercader and Levy, 2004). 4. Funtional Data Analysis and functional ANOVA The first steps of Statistics were characterized by the analysis of random univariate data. Even in regression models, the response variable and the regressor were originally one dimensional objects. Ramsay and Silverman (1997) express it saying that at the beginning the statistical atom was a number (an element of R). The collection of statistical techniques that form introductory courses in Statistics are designed to analyzed this kind of data. In the first half of the twenty century Statistics reached maturity and started to deal with multivariate data: a collection of p random numbers observed simultaneously (in the same subject) was treated as a statistical atom (an element of p R ). Principal component analysis, discriminant analysis, and cluster analysis, among others, are part of what we know today as Multivariate Data Analysis. At the end of the last century a new statistical object has started to be considered. In the last years, the joint development of real-time measurement instruments and data storage computer resources has made possible to observe and save complete functions as results of random experiments. For instance, continuous-time monitoring clinical diagnostics or stock market information are common nowadays. From this perspective, random functions are now the statistical atoms. Functional Data Analysis (FDA) deals with the statistical description and modelization of samples of random functions. It’s well worthwhile noting that random functions can also be obtained from standard random samples, by the application of nonparametric curve estimation methods. For instance, Kneip and Utikal (2001) study the temporal evolution of density function of incomes in United Kingdom from 1968 to 1988. They work with yearly cross-sectional sample of households, and use nonparametric density estimation methods (kernel methods, to be specific) to obtain 21 income density functions over time, one corresponding to each year. Other kind of data can be transformed into functions using nonparametric regression (also known as smoothing methods). 7 A broad outlook to Functional Data Analysis is given in the books by Ramsay and Silverman (1997, 2002). They show interesting applications and case studies and include many of the available techniques. Most of them are versions of standard statistical methods adapted to functional data (for instance functional descriptive statistics, functional linear models or functional principal component analysis; see also Cuevas et al, 2004, for more references on these topics) but others are specific for this kind of data because they exploit the nature of functions: for example, principal differential analysis is a kind of principal component analysis made on the derivatives of the observed functions, and regularization is a pre-process step where a change of variable is done in each observed function in order to made them as similar as possible. One of the standard methods that have been generalized to be used in Functional Data Analysis is the one-way analysis of variance (functional ANOVA). In Ramsay and Silverman (1997) it is included in what they call functional linear models (a generalization of linear regression models). In addition to this approach, there are works devoted specifically to this topic, among the ones we highlight Cuevas et al. (2004) because it is the base of our work. In functional ANOVA it is assumed that the n observed functions )(xfij follow the model ),()()( xexmxf ijiij += [ ] ,,,,,1,,,1 baxkinj i ∈ = = KK ,nn ii= ∑ where )(xmi are k unknown mean functions and )(xeij are independent trajectories drawn from a 2 L-process with zero mean and covariance function ),())(),(( yxKyexeCov ijij =. So the observations )(xfij constitute k independent samples of random functions, each with a specific function mean and a common covariance structure. This is the homocedastic case. The heteroscedastic version allows for a different covariance function in each sample: ),())(),(( yxKyexeCov iijij =. The hypothesis to be tested in both cases is ).()(: 10 xmxmH k = = L As stated, in the context of the present paper, )(xfij is the relative equivalent income (before or after taxes and transfers) estimated density function of region j in country i, one of the 15=k countries forming the European Union before May 2004. So the data x corresponding to a person represents that its household has a relative income equal to x times the median regional equivalent income. The interval [ ] ba,, where values x belongs to, is [] 7,1− when incomes before taxes are transfers are used, and it is [] 5,1− for data after taxes and transfers. Section 5 deals with the estimation of density functions. The null hypothesis establishes that regional income densities have the same mean value in each country. Another way of wording it is to say that under the null hypothesis there is no country effect in the observed variability of regional income densities. The classical F statistic for the univariate one-way ANOVA computes the ratio of variabilities between samples and intra sample. The functional version, as stated in Cuevas et al. (2004), would be 8 ∑ ∑ −− −− = • =••• ji iij k iii nknff kffn F , 2 1 2 )/( )1/( , where )(tff •••• = is the global mean function, )(tff ii •• = is the mean function in the i-th sample, and 2/1 2)(       =∫b adxxff is the usual 2 L norm. The null hypothesis 0 H should be rejected if the numerator of n F (a measure of the differences between groups) is to big, compared with the denominator of n F (a measure of the variability of the noise process generating )(xeij ). Cuevas et al. (2004) argue that it is enough to consider the numerator of n F when you are comparing values of the statistic coming from functional ANOVA models with noise processes having the same variability (all the denominators are estimating the same quantity). This is the case when an observed n F value is compared with Monte Carlo simulated values and the simulation is done to produce data according to the null hypothesis and having the same noise variability as the observed data. Technical reasons lead Cuevas et al. (2004) to measure differences between groups using the statistic ∑ < •• −= ji jiin ffnV 2, that is equivalent to use the numerator of n F. Their Theorem 1 establishes that the asymptotic distribution of n V under 0 H coincides with that of the statistic ∑ < −= ji jiji ZCZV 2, where 2/1 )/( jiij ppC =, ii pnn →)/( as ∞ →n, and kixZZ ii ,,1),( K= = , are independent Gaussian processes with 0 mean and covariance function ),( yxKi, that can be consistently estimated by ()() )()()()( 1 1 ),( ˆ 1 yXyXxXxX n yxK iij n j iij i i i • = •−− − =∑. In the homoscedastic case the natural estimator of the common covariance function is ∑ = − − =i n j ii yxKn n yxK 1 ),( ˆ )1( 1 1 ),( ˆ. This theoretical result offers an asymptotic Monte Carlo procedure to tabulate the null distribution of n V: a large number N of values of statistic Vare simulated (say NlVl,,1, *K=) and the p-value corresponding to the observed n V is computed as the proportion of simulated values * l V greater than n V. Assuming homoscedasticity, an alternative way to obtain valid approximate the null distribution of n V is based on permutations. The procedure to obtain pseudo functional data sets according to the null hypothesis consists in randomly permute the group label of the observed functions. Under the null hypothesis the original sample and the permutated sample are interchangeable, and so they are the observed n V and the value 9 p Vcomputed in the permutated sample. N values of statistic p Vare obtained by permutation and the p-value corresponding to the observed n V is computed as the proportion of times the permuted statistics p V is greater than n V. Also in the context of functional ANOVA, Muñoz-Maldonado et al. (2002) use a permutation test based on a different statistic. The permutation test presents a drawback under the alternative hypothesis. The between group variability is translated by the permutation procedure to noise variability in the permuted samples. Therefore the artificial samples verify the null hypothesis of equal mean in different groups, but the noise variability is greater than the corresponding to the original data. The main consequence of the increment in noise variability is the reduction of the test power: small deviation from the null hypothesis would not be detected because of the precision loss. The preceding procedure can be modified to obtain a more powerful permutation test. Instead of permuting the observed function, we propose to permute the estimated residuals and define the artificial functions as the sum of the global mean plus a permutated estimated residual, )( ˆ )()( xexfxf p ij p ij += •• , and )( ˆxe p ij is selected from the estimated residuals, • −= rrlrl fxfxe )()( ˆ, by random permutation. When the null hypothesis is false this modified permutation test guarantees that the artificial data verifies the null hypothesis and have approximately the same noise variability as the original sample. Therefore this modified test would detect deviations from the null hypothesis that would be overlooked by the standard permutation test. From now on we use this modified permutation test when we refer to permutation test in text. Let us now discuss an important question arising in economic and social micro data bases: in theses contexts it is usual each observation has a different weight, that is proportional to the amount of people in the population it is representing. The functional ANOVA test based on the statistic n V, as in Cuevas et al. (2004), is not directly applicable to such weighted samples. Delicado (2006) establishes for weighted samples the result analogue to Theorem 1 in Cuevas et al. (2004). A new statistic w n V, the version of n V appropriate for weighted samples, is introduced. Its null distribution can be approximated by Monte Carlo simulation or by permutation methods. The procedures described here can be used to test the null hypothesis of no factor effects in the functional ANOVA model for any k samples of functions in 2 L. If we are comparing samples of density functions (as it is the case in this work), that are always in 1 L, we can assume that they also belong to 2 L and use the above methods. Moreover a 1 L version of the n F statistic can be defined as ∑ ∑ −− −− = • =••• ji iij k iii L nknff kffn F ,1 11 )/( )1/( 1, (1) where the norm 1 L of a function f is defined as ∫ =b adxxff )( 1. 16 corresponding to the net income densities, what corroborates our thesis on country effects. More formal support can be obtained computing distances of 1,T d and 2,T d kind. After Before 0 5 10 15 L2; Assymptotic Heteroscedastic After Before 51015 L2; Assymptotic Homoscedastic After Before 51015 L2; Permutation test After Before 0.02 0.04 0.06 0.08 L1; Permutation test Figure 5. Box-plots of the N simulated (or permuted) values of distributions of statistics w n V and 1 L n F. Observed values of statistics w n V and 1 L n F as big solid circles. The null hypothesis of no country effect is rejected in all cases: equivalent income before or after taxes and benefits, w n V or 1 L n F, and the three ways to approximate the null distribution of w n V. In fact, the 8 resulting p-values have been estimated as 0: no simulated or permuted sample of functions presents statistics w n V or 1 L n F greater than those observed in our original functional data set, implying that distances 1, w n V d and 1, 1 L n F d, based on p-values, are always equal to 1, for both sets of income densities (before and after taxes and benefits). Therefore this kind of distances between data and null hypothesis does not help to validate our thesis that country effect is more important after taxes and benefits. 17 When computing type 2,T d measures, based on the standardized observed statistic value, things are different and it is possible to calibrate the dissimilar behaviour of regional densities before and after the action of tax-benefit systems. Mean and standard deviations under the null hypothesis are estimated from simulated or permuted samples. Table 4 shows distances 2, w n V d (columns 1 to 3) and 2, 1 L n F d (column 4) for situations before (first row) and after (second row) taxes and benefits. All those numbers indicate that the evidence against the null hypothesis of no country effects after is more than twice than before taxes and benefits (see the third row, what shows the quotient of rows 1 and 2). Results corresponding to asymptotic approximations (columns 1 and 2) are consistent, as well as they are those based on permutation tests (columns 3 and 4). The similarity between the two first columns indicates that homoscedasticity is a valid assumption. Discrepancies between these columns and the last two lead to question the soundness of asymptotic approximations: in fact asymptotic approximations need a large number of functions per class (large number of regions per country, here) what does no happen in our case. Despite those discrepancies, values in the third row are similar for the four ways to carry on the functional ANOVA tests, what points out the robustness of our conclusion on the used method. Equivalent income density... w n VT = Heteroscedastic asymptotic approximation w n VT = Homoscedastic asymptotic approximation w n VT = Permutation test approximation 1 L n FT = Permutation test approximation ... before taxes and benefits 15.03 10.98 5.20 6.20 ... after taxes and benefits 36.34 34.22 12.84 14.55 Quotient of the previous rows 2.42 3.12 2.46 2.35 Table 4. List of distances 2,T d for four versions of the functional ANOVA test, and two scenarios. 7. Concluding comments In this paper we have studied the role of the country factor in explaining the variability of regional income distributions before and after national tax-benefit systems operate in Europe-15. Our main aim has been to disentangle the relative contribution of countries’ tax-benefit systems on income distribution shapes, in relation to other country-specific institutional features. Functional Data Analysis techniques have allowed to take into account of the full income distribution, instead of a single univariate index. Moreover, the Functional ANOVA model has permitted the analysis of the country explicative power on the dispersion of regional income density functions (estimated nonparametrically) before and after taxes and transfers. Our statistical evidence shows that regional income distributions of different countries are statistically different, both before and after redistribution takes place –we reject in both cases the null hypothesis of equal average regional distributions among countries. 18 The analysis of the full distribution leaves therefore to different conclusions than the analysis based on a single inequality index –remember that using the Gini we did not reject the null hypothesis of equality of the Ginis on the distribution before. However, the null assumption is rejected more strongly in the after redistribution case. If we consider the value of the statistic used in the contrast as a distance between the observed data and the model under the null hypothesis, the distance separating (from equality) the regional distributions after is more than the double (and smaller than tree times) the distance before. From a European perspective, our results suggest a different reading. The variability between countries in the shape of their regional income distributions is clearly increased by national tax-benefit systems. Our analysis suggests that a European taxes and benefits would contribute reducing disparities in disposable income distribution among countries. 19 Bibliography Atkinson, A.B. (2000) “Increased Income Inequality in OECD Countries and the Redistributive Impact of the Government Budget”, WIDER, World Institute for Development Economics Research, Working Papers No. 202, October. Atkinson, A. B. Bourguignon, F. and Chiappori, P.A. (1988) « Fiscalité et transferts: une comparaison franco-britannique ». Annales d'Economie et de Statistique, v. 0, iss. 11, 117-40. Atkinson, A.B. Rainwater, L. and Smeeding, T.M. (1995), Income Distribution in OECD countries. Evidence from the Luxembourg Income Study, Social Policy Studies nº 18, OECD, Paris. Bowman, A.W. and Azzalini, A. (2001) Applied Smoothing Techniques for Data Analysis. Oxford. Oxford University Press. Cuevas, A., Febrero, M., and Fraiman, R. (2004) An anova test for functional data. Computational Statistics & Data Analysis, 47, 111-122. Delicado, P. (2006) Functional ANOVA when data are a weighted sample of density functions. Preprint. Jenkins, S.P. and Van Kerm, P. (2005) Accounting for income distribution trends: a density function decomposition approach. Journal of Economic Inequality, 3, 43-61. Kneip, A., and Utikal K.J. (2001) Inference for Density Families Using Functional Principal Component Analysis. Journal of the American Statistical Association, 96, 519-542. Locantore, N., Marron, J.S., Simpson, D.G., Tripoli, N., Zhang, J.T., and Cohen, K.L. (1999) Robust principal component análisis for functional data. TEST, 8 (1), 1-73. Mercader, M. and Levy, H. (2004) The role of tax and transfers in reducing personal Income Inequality in Europe’s regions: Evidence from EUROMOD. EUROMOD Working Paper No. EM9/04. Muñoz-Maldonado, Y., Staniswalis, J.G., Irwin, L.N., and Byers, D. (2002) A similarity analysis of curves. Canadian Journal of Statistics, 30, 373-381. R Development Core Team (2005). R: A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria. ISBN 3-900051-07-0, http://www.R-project.org. Ramsay, J.O., and Silverman, B. W. (1997) Functional Data Analysis. New York: Springer. Ramsay, J.O., and Silverman, B. W. (2002) Applied Functional Data Analysis: Methods and Case Studies. New York: Springer-Verlag. Sheather and Jones (1991). A reliable data-based bandwidth selection method for kernel density estimation. Journal of the Royal Statistical Society, Series B, 53, 683-690. Stewart, K. (2002) “Measuring Well-Being and Exclusion in Europe’s Regions”, CASEpaper 53, CASE Centre for Analysis of Social Exclusion, London School of Economics. Sutherland, H. (Ed.) (2001) “Final report. EUROMOD: an integrated European benefittax model”, EUROMOD Working Paper nº. EM9/01. Silverman, B.W. (1986) Density Estimation for Statistics and Data Analysis. London: Chapman and Hall. Wagstaff et 24 other authors (1999). “The redistributive effect, progressivity and differential tax treatment: Personal income taxes in twelve OECD countries”, Journal of Public Economics, 72, 73-98. Wand, M.P. and Jones, M.C. (1995) Kernel Smoothing. London: Chapman and Hall. 20 Table 1: Databases used in each country Country Base Dataset for EUROMOD Date of collection Reference time period for incomes Austria Austrian version of European Community Household Panel (W5) 1999 annual 1998 Belgium Panel Survey on Belgian Households (W6) 1999 annual 1998 Denmark European Community Household Panel (W2) 1995 annual 1994 Finland Income distribution survey 1998 annual 1998 France Budget de Famille 1994/5 annual 1993/4 Germany German Socio-Economic Panel (W15) 1998 annual 1997 Greece European Community Household Panel (W2) 1995 annual 1995 Ireland Living in Ireland Survey (W1) 1994 month in 1994 Italy Survey of Households Income and Wealth 1996 annual 1995 Luxembourg PSELL-2 (W5) 1999 annual 1998 Netherlands Sociaal-economisch panelonderzoek (W3) 1996 annual 1995 Portugal European Community Household Panel (W3) 1996 annual 1995 Spain European Community Household Panel (W3) 1996 annual 1995 Sweden Income distribution survey 1997 annual 1997 UK Family Expenditure Survey 1995/6 monthly in 1995/6 Table 2. Regional information in EUROMOD national databases Country Information used in our analysis Average population per region Number of regions Austria NUTS1 2,645,969 3 Belgium NUTS1 3,299,186 3 Denmark Not considered Finland NUTS2 814,532 6 France NUTS2 2,588,033 22 Germany NUTS1 4,966,341 16 Greece NUTS1 2,635,525 4 Ireland Not considered Italy NUTS1+ South Split 4,767,237 12 Luxembourg Not considered N etherlands Not considered Portugal NUTS2 1,417,429 7 Spain NUTS1 5,557,837 7 Sweden NUTS2 1,123,557 8 UK NUTS1+Greater London 4,786,980 12 21 Table 3: Regions in EUROMOD and sample sizes. Country Country Code Region Name NUTS abbreviation Sample size Portugal 12 Madeira PT3 598 12 Açores PT2 599 12 Algarve PT15 637 12 Centro PT12 1,027 12 Alentejo PT14 514 12 Norte PT11 840 12 Lisboa E Vale Do Tejo PT13 591 Greece 7 Thraki, Makedonia, Thessalia GR1 1,660 7 Dellada, Sterea, Pelloponisos, Ionia Nisia Ipiros GR2 1,251 7 Notio Aigaio, Voreio Aigaio, Kriti GR4 656 7 Athina GR3 1,601 Spain 13 Canarias ES7 380 13 Andalucia, Murcia ES6 1,013 13 Cast Leon, Cast Mancha, Extremadura ES4 959 13 Galicia, Asturias, Cantabria ES1 895 13 Catalunya, Valen, Baleares ES5 1,375 13 Euskadi, Navarra, Rioja, Aragón ES2 960 13 Madrid ES3 537 Italy 9 Sicilia ITA 559 9 Basilicata/Calabria IT92 389 9 Campania IT8 709 9 Puglia IT91 520 9 Sardegna ITB 295 9 Abruzzo-Molise IT7 396 9 Lazio IT6 411 9 Center IT5 1,250 9 North-east IT3 1,009 9 North-west IT1 1,048 9 Lombardia IT2 824 9 Emilia IT4 725 Belgium 2 Flandre BE2 1,961 2 Wallonie BE3 1,300 2 Bruxelles BE1 393 Sweden 14 Smaland med oarna SE03 1,607 14 Norra Melansverige SE06 1,235 14 Mellersta Norrland SE07 888 14 Ovre Norrland SE08 1,049 14 Vastsverige SE05 4,448 14 Sydsverige SE04 2,889 14 Östra mellansverige SE02 3,404 14 Stockholm SE01 4,114 Finland 4 Itä-suomi FI13 1,394 4 Väli-suomi FI14 1,364 4 Pohjois-suomi FI15 762 4 Etelä-suomi FI12 3,276 4 Ahvenanmaa/aland FI2 44 4 Uusimaa FI11 2,152 UK 15 Northern Ireland UKB 134 15 North UK1 405 15 West Midlands UK7 621 22 15 Scotland UKA 604 15 Yorks & Humberside UK2 594 15 East Anglia UK4 282 15 North West UK8 722 15 Wales UK9 339 15 East Midlands UK3 491 15 South West UK6 637 15 South East UK5 1,274 15 Greater London UK55 694 Germany 6 Thüringen DEG 358 6 Sachsen-Anhalt DEE 356 6 Sachsen DED 594 6 Mecklenburg-Vorpommern DE8 216 6 Brandenburg DE4 318 6 Rheinland Pfalz / Saarland DEB 417 6 Bremen DE5 61 6 Berlin-Ost DE3 189 6 Niedersachsen DE9 626 6 Baden-Würtemberg DE1 934 6 Hamburg DE6 97 6 Bayern DE2 984 6 Saarland DEC 134 6 Nordrhein Westfalen DEA 1,508 6 Schleswig-Holstein DEF 190 6 Hessen DE7 498 France 5 Corse FR83 38 5 Nord-Pas De Calais FR3 715 5 Basse-Normandie FR25 281 5 Auvergne FR72 234 5 Champagne-Ardennes FR21 305 5 Poitou-Charentes FR53 309 5 Languedoc-Roussillon FR81 437 5 Bretagne FR52 582 5 Pays De La Loire FR51 653 5 Centre FR24 434 5 Limousin FR63 153 5 Franche Comte FR43 256 5 Bourgogne FR26 320 5 Aquitaine FR61 585 5 Midi-Pyrenees FR62 523 5 Haute-Normandie FR23 355 5 Picardie FR22 305 5 Rhone-Alpes FR71 949 5 Provence-Alpes-Cote Dazur FR82 879 5 Lorraine FR41 472 5 Alsace FR42 367 5 Ile De France FR1 2,139 Austria 1 West: Oberösterreich, Salzburg, Tirol, Vorarlberg AT3 879 1 South: Kärnten, Steiermark AT2 648 1 East: Wien, Burgenland, Niederösterreich AT1 1,145