Social Investment and Youth Labor Market Participation

In this paper, we first rely on small area techniques to derive from EU statistics on income and living conditions (EU&#8208;SILC) survey new indicators of compensatory and social&#8208;investment policies at regional level. While compensatory policies have mainly the goal of protecting individuals from &#8220;old&#8221; risks (e.g., old&#8208;age), investment&#8208;related social policies tend to focus more on &#8220;new social risks&#8221; (e.g., skill deficits). We rely on these new indicators to perform a data&#8208;driven structural vector autoregressive (SVAR) analysis to investigate the causal relationships between youth labor market outcomes and these two types of spending. Our results support the view that social&#8208;investment policies are effective for tackling new social challenges.


I. INTRODUCTION
Since its inception, the EU has experienced robust convergence in terms of gross domestic product (GDP) per capita. However, even though there was a convergence process at the country level, the convergence at the regional level has been much weaker. In particular, there are still some countries exhibiting regional divergence or sustained North-South (or West-East) divides (Monfort 2008;Wunsch 2013). This means, for example, that there tends to be much higher negative correlation between GDP and unemployment within countries than across countries. However, both mainstream and heterodox theories cannot explain the existence of these different regional trajectories and the weakness of the convergence processes among them (Iammarino, Rodríguez-Pose, and Storper 2018).
Importantly, there is still considerable crosscountry variation when it comes to youth labor * The author would like to thank Kevin Albertson, Marco Capasso, and Alessio Moneta for useful comments and suggestions, and Alessio Moneta for providing R codes. This project has received funding from the European Union's Horizon 2020 research and innovation program under grant agreement No 649189, "Innovative Social Investment: Strengthening communities in Europe" (InnoSI).
In lovely memory of a young European, Antonio. European countries (especially Germany and Austria) have a lower youth unemployment and inactivity rates, along with higher employment rates, than the rest of the EU-especially Southern and Eastern Europe (Pastore 2018;Tomić 2018). The recent financial crisis has further deepened regional disparities having a profound impact on the employability of the young (Bruno, Marelli, andSignorelli 2014, 2015). In this regard, the worst changes have been recorded in Southern regions, where the youth unemployment rate has doubled or tripled since the onset of the recession (Bruno, Marelli, and Signorelli 2014;Mascherini et al. 2012). Chen et al. (2018) also show that the risk of poverty for the young (and the working-age population) has increased significantly since the financial crisis in 2008, while it has declined sharply for the elderly.
The main contribution of this paper is thus to empirically investigate-at a regional level-the role that different policies adopted in the recent years had on youth employability. Except from few studies (Bruno, Marelli, and Signorelli 2015), there are limited regional studies on that topic. Traditional indicators of labor market participation, however, such as unemployment and youth employment rates, do not adequately capture new "gray" area that represent market attachment in contemporary societies (Mascherini et al. 2012). For this reason, in addition to traditional indicators of young employment and unemployment rate, in this paper, we also focus our attention on the share of young people that are disengaged from both work and education, usually indicated with the term NEETs (not in employment, education, and training). The needs to focus more on NEETs is now central in the European policy debate, and the term is explicitly mentioned in the Europe 2020 agenda as well as in the 2012 Employment Package "Towards a job-rich recovery" (Eurofond 2012).
In particular, in line with the recent literature on social investments (Hemerijck 2013;Nikolai 2012), our aim is to differentiate between two broad types of policies: social-investment and compensatory policies. Compensatory policies are mainly based on a contribution-financed social security with the goal of protecting individuals from "old" risks, such as unemployment and old-age. Social-investment policies tend to focus more on "new social risks" to overcome, through education and training, skill deficits that may emerge in postindustrial labor markets (Nikolai 2012). Furthermore, these policies tend to reconcile work and family life. Thus, the focus is on investment in human capital as well as the provisions for the needs and the future of the younger generations. Indeed, several studies have documented a transition from the traditional welfare state to a new investment state in many European countries (Bonoli 2007;Ferragina, Seeleib-Kaiser, and Spreckelsen 2015;Obinger and Starke 2014). For example, Nikolai (2012) finds mixed evidence in support of a shift toward more social investment, with Continental and Southern European Countries being characterized by more spending for compensatory and less spending for investment-related policy (especially education).
However, it is quite impossible to properly assess the impact of the two types of policies without having expenditure data disaggregated at a regional level. As it has also been highlighted by the DG Regional Policy of the European Commission, in order to better target policy measures, there is an increasing need of social policy indicators developed at regional regional level (Commission 2010;Verma, Gagliardi, and Ferretti 2013). Therefore, the second contribution of our paper is to present new indicators of regional spending (which are comparable across regions and countries) which are derived through the cumulation methodology applied to the EU statistics on income and living conditions (EU-SILC) dataset (Betti et al. 2012). In particular, we are able to compute-for a subset of European countries-the average amount of cash transfers that a household received for each category of compensatory and social-investment spending in a year. In doing so, we are thus able to derive for each category of spending a regional indicator which is comparable across time and across countries, and tends to be-compared to national measures-more precise for monitoring and assessing the effectiveness of each policy. Indeed, the recent "Youth Guarantee" program targeting all the young people under 25 years in Europe is implemented at a regional level. Moreover, these regional indicators will allow us to (indirectly) take into account the important role of intrafamily transfers as suggested by recent studies (Gál, Vanhuysse, andVargha 2018, Francesconi andHeckman 2016).
We then investigate the impact of these indicators on youth labor market participation within a structural vector autoregressive (SVAR) framework. In particular, we rely on a data-driven approach, recently introduced in the literature by Moneta et al. (2013), which rely on independent component analysis to identify structural parameters in SVAR (Gouriéroux, Monfort, and Renne 2017;Lanne, Meitz, and Saikkonen 2017;Shimizu et al. 2006). Specifically, we adopt an identification scheme, called LiNGAM, that is, linear non-Gaussian acyclic model (Shimizu et al. 2006), to identify contemporaneous parameters in order to describe the causal relationships among variables. Differently from standard methods (such as Cholesky decompositon), which necessarily requires either a theoretical justification or an institutional knowledge, this method has the great advantage to achieve identification of structural parameters directly from the data and statistical analysis alone. More specifically, this method allows us to identify the exogenous shocks affecting-at each time and in an independent way-our policy variables, that is, the level of compensatory and social-investment spending. In doing so, we are thus able to identify-through the impulse-response function (IRF)-both the direct and indirect effects that a shock in our policy variables has on youth employment outcomes. For example, we will be able to assess how a shock in the level of (household) compensatory spending directly impacts on the level of youth unemployment, as well as how it will indirectly affect (either through family links or an increase in the level of GDP per capita) the level of NEETs.
Our analysis of regional spending suggests that, even though the evidence is consistent with previous analyses using national data to what concern the compensatory component (Hemerijck 2013(Hemerijck , 2017Heitzmann, Wukovitsch, et al. 2015), there is higher regional variation in the investment component, even within the same country. The results from our SVAR analysis also suggest that investment policies are effective to reduce the level of NEETs and increase the level of youth employment.
In the following, we first give a brief overview of the literature on youth labor market participation. We then describe how we derive our dataset. In particular, in Section III, we briefly review the main statistics on labor market participation of the young, which are currently available at Eurostat, and the main issues related to regional data on expenditure. In Section IV, we describe the cumulation methodology, and we apply it to EU-SILC in order to develop indicators of compensatory and social-investment spending at regional level, while in Section V we rely on a recently econometric methodology developed by Moneta et al. 2013 to investigate the effects of these types of policies on labor market outcomes. Section VI concludes our argument.

II. LITERATURE REVIEW
The number of regional studies about the impact of recent policies on youth labor market participation is still rather limited. Moreover, very few of them address the issue of NEETs specifically. Indeed, while NEETs and youth (un)employment are related concepts, there are important differences. In particular, unemployment rate measures the share of the labor population who are not able to find a job. More precisely, it is a measure of those who are out of work, but have actively looked for work in the recent past and is available for work in the near future. However, this measure does not take into account the "new risks," that is it does not capture those who became discouraged and decided to stop looking for a job (Eurofond 2012;Mascherini et al. 2012).
This implies that the unemployment rate may stop falling even when a relevant number of individuals are at high risk of labor market and social exclusion. A similar remark can be made for youth employment rate, which measure the share of the working-age population (i.e., people aged 15 to 24) who is currently employed. In contrast, the NEETs-as defined by the European Commission (DG EMPL)-captures the share of the young population currently disengaged from the labor market and education, namely unemployed and inactive young people not in education or training. 1 In particular, a number of recent papers confirm a larger impact of the recent financial crisis on youth employment rates when compared to adult rates, as well as a greater responsiveness of young employment to the business cycle (Bruno, Marelli, and Signorelli 2015;Coppola and O'Higgins 2015;O'Higgins 2012). More precisely, when an economy is expanding, its youth unemployment rate decreases more than the average, while it increases more than average when an economy is contracting (Pastore 2018). The main reason have been identified in the lower qualifications of the young, their experience "gap" along with weaker work contracts (Tomić 2018). Another important factor is the role of educational policy and differences in school-to-work transition regimes (Corsini and Brunetti 2018). O'Higgins (2012) also highlights that the negative effects of a recession are likely to last longer for the young.
Very few papers, though, analyze the effect of various policies on NEETs by comparing different European countries. Bruno, Marelli, and Signorelli (2015) show that NEET rates, both for male and female, are persistent over time to a degree comparable to youth unemployment rate, and this persistence further increases over the crisis years. Moreover, the sensitivity of NEETs to GDP substantially decreases: an increase in GDP An alternative measure is to look at the ratio of youth to adult unemployment rates, which more closely reflects a country's institutional characteristics and the functionality of its school-to-work transition system (Corsini and Brunetti 2018;Pastore 2018). after many years of stagnation would thus only have a limited impact on the situation of young people. Thus, the policy implication of this analysis is that both structural and counter-cyclical policies are required. However, no direct indicators for the policies at regional/national level are available. Finally, Caliendo and Schmidl (2016) provide a comprehensive survey of the recent evidence on the effectiveness of active labor market programs (ALMP) for youth in Europe. Overall, the aggregate evidence is somewhat discouraging: while job search assistance results in positive effects for young employability, there are clearly negative effects for public programs and mixed evidence for training and wage subsidies.
To sum up, recent studies on labor market participation of the young in Europe suggest that young employability has significantly worsen in recent years, especially in Eastern and Southern counties. Importantly, (un)employment and NEETs rates do not quickly respond to an increase in GDP. However, there is no studies which specifically address-at a regional level-the role of compensatory and investment policies in tackling this issue. That will be the objective of the following empirical analysis.

III. ISSUES WITH REGIONAL DATA
In this section, we describe the economic indicators of youth labor market participation we will use in our analysis (i.e., our outcome variables), and we discuss the main issues related to the collection of regional data on expenditure (i.e., our policy variables).

A. Regional Data on Young People's Labor Market Participation
We report measures for unemployment, employment and NEETs rate for the 15-24 age group in Table 1, Table 2, and Table 3 as computed at NUTS1 level.
In particular, these tables report for each variable, in addition to the mean (μ) and the standard deviation (σ) computed at country level, the coefficient of variation (CV). This latter indicator is a normalized measure of dispersion defined as the ratio between the standard deviation and the mean (i.e., σ |μ| ). For a given standard deviation value, it thus indicates a high or low degree of variability only in relation to the mean value. Since the CV is a measure of relative variability which is unit-free (i.e., does not depend on the unit of measurement), it is often preferred to the standard deviation which has no interpretable meaning on its own. In particular, the CV indicators is among those indicators of σ − convergence, which is a term used to refer to a reduction of disparities among regions over time (Monfort 2008). 2 For example, from Table 1, we can observe that high level of youth employment rates can be observed in Austria (AT), Denmark (DK), Finland (FI), the Netherlands (NL), and United Kingdom (UK). Conversely, young people seem particularly disengaged from the labor market in Slovakia (SK), Bulgaria (BG), Lithuania (LT), Italy (IT), Hungary (HU), and Greece (GR). Moreover, although there is a moderate variation in youth employment rate across European countries, there is a greater variation in youth unemployment rate (with the CV being as high as 50%, see Table 2). The level of NEETs is also very different among EU countries, although slightly lower than unemployment rates (see Table 3). However, once again Southern and Eastern Europe counties tend to have the higher NEET rates.
In these tables, we also focus our attention on the time variation that occurred in our outcome variables between years 2009-2007 and 2013-2011. These measures will be specifically used in our empirical analysis (see the next section), and highlight a huge impact of the financial crisis on youth employability. In line with Bruno, Marelli, and Signorelli (2015), it is possible to notice the dramatic variation in unemployment rates that occurred in Eastern and Southern countries, being around around +20% in 2009-2007. The variations in NEETs rates have been smaller (at most about 6%). In that case, though, it is possible to notice that Anglo-saxon countries performed worse compared to new Member States. In addition, these tables highlight a significant variation in country responsiveness to the financial crisis. While in some countries, the situation of the young experienced a recover in 2013-2011, in other countries, especially the Southern European ones, experienced a further worsening.
Finally, as Figure 1 suggests, the EU-28 CV computed at NUTS1 level is increasing over time for all these measures. This suggests a divergence among EU countries in the level of unemployment, employment, and NEETs. Notes: The employment rate is computed as the share of employed young over the working population (15-24 years old). Statistics are computed relying on data available at Eurostat for years 2007-2013. The coefficient of variation (CV) is computed by dividing the country mean over the standard deviation. The regional coefficient of variation (Reg. CV) is computed by dividing the regional mean over the regional standard deviation (where NUTS1 data are available).
It is important to notice, that the increase in Regional disparities within EU as a whole does not prevent disparities from decreasing within each Member states (Monfort 2008). For this reason, we also compute CV indicators for each Member State at regional level (where NUTS1 level data are available). However, even when we look at the regional variation within countries for the same variable, we can notice that for some countries, the regional variation can be very large: for example, in Italy and Portugal the CV is about 40%.

B. Regional Data on Expenditure
Social policies that are defined as socialinvestment policies are usually categorized according to three aspects they promote (Heitzmann, Wukovitsch, et al. 2015): 1. Maintenance or restoration of the capacity of labor market participants (e.g., old age pensions).
2. Entrance of new labor market participants (short-term unemployment insurance; short-term maternity leave).
3. Investment in the capacity of new labor market participants (elderly care, child care).
Unfortunately data on these dimensions are often not available at regional level and for several years. For these reasons, any attempt to examine the development of social investment across regions and countries often fails. Even if alternative approaches are available (De Deken 2014), because of data limitation, researchers largely end up with two categories, one for compensatory (i.e., the old risk categories) and another for social-investment policies (i.e., the new risk categories).

Notes:
The unemployment rate is computed as the share of unemployed young over the labor force (15-24 years old). Statistics are computed relying on data available at Eurostat for years 2007-2013. The coefficient of variation (CV) is computed by dividing the country mean over the standard deviation. The regional coefficient of variation (Reg. CV) is computed by dividing the regional mean over the regional standard deviation (where NUTS1 data are available).
In this analysis, we similarly distinguish between these two broad categories, but in addition to previous research, we rely on data from EU-SILC survey to derive indicators at country regional level. The EU-SILC is a very rich survey on income and social condition collected at household (and individual) level under a standard integrated design by nearly all EU countries. As explained below, we rely on small area estimation (SAE) techniques-in particular on cumulation technique-to derive regional indicators of investment and compensatory policies from EU-SILC survey (Betti et al. 2012;Verma, Gagliardi, and Ferretti 2013). Specifically, for each category of spending (social-investment and compensatory), we derive a series of indicators by computing the average amount received per household at NUTS1 level. This an important contribution to previous studies, in which indicators of total spending where usually derived-at a country level-as a share of the GDP (Prandini, Orlandini, and Guerra 2016 on this issue).

IV. CUMULATION METHODOLOGY AND EU-SILC VARIABLE SELECTION
EU-wide comparative datasets such as EU-SILC, even though primarily developed to construct indicators at the national level, can serve as a unique source for generating comparative indicators at regional levels through SAE techniques. Such methodologies have already been proved to be successful to derive regional measures of poverty (Betti et al. 2012;Marchetti et al. 2015;Verma, Betti, and Gagliardi 2010;Verma, Gagliardi, and Ferretti 2013).
In particular, we rely on (average) measures, which are obtained by cumulating and consolidating the information over waves of national

Notes:
The NEET rate is computed as the share of young not education, employment and training over the working population (15-24 years old). Statistics are computed relying on data available at Eurostat for years 2007-2013. The coefficient of variation (CV) is computed by dividing the country mean over the standard deviation. The regional coefficient of variation (Reg. CV) is computed by dividing the regional mean over the regional standard deviation (where NUTS1 data are available). sample surveys in order to obtain indicators that permit greater spatial disaggregation (Verma, Gagliardi, and Ferretti 2013). To be able to compute spatial statistics through cumulation, the only information required is the strata identifiers from which individuals are sampled from. More specifically in our case, to cumulate over waves we need to know from which NUTS1 region the individuals were sampled. Unfortunately, in our sample, this information is only available for a limited numbers of countries, namely: Austria, Belgium, Bulgaria, Czech Republic, France, Greece, Hungary, Italy, Poland, Spain, Sweden, and United Kingdom. Therefore, only for this group of countries, we are able to derive an indicator of regional spending at NUTS1 level along with a measure of dispersion (i.e., the regional CV). For the remaining group of countries, we can only derive country-level indicators from EU-SILC.
We proceed as follows. Given that we have the cross-sectional dataset of the EU-SILC survey for nine consecutive years (from 2006 to 2014), the objective is to compute the cumulative average of a given measure y over 3 years, that is, y c t . We first construct for each year (i.e., for each EU-SILC wave) the yearly average relying on N individual observations (i.e., y t = 1 N ∑ N i=1 y i ). Then for each year t, we estimate the required statistic y c t as the 1-year moving average over three consecutive years of the annual average y t , that is However, to allow for more variability in our dataset, we only allow for one overlapping year across observations, relying therefore on four central years, that is, we select y  In particular, we rely on the following EU-SILC variables to derive the level of compensatory spending (we report the EU-SILC number and a detailed description for each variable in the online Appendix): 1. Unemployment benefits.

Old-age and survivor benefits. 3. Sickness benefits. 4. Disability benefits.
Similarly to derive the level of socialinvestment policies, we select the following variables: 1. Education-related allowances. 2. Family/children allowances. 3. Housing allowance.
More generally, both groups of variables are defined as current transfers received by the household during the reference period, through collectively organized schemes, or outside such schemes by government units and nonprofits institutions serving households (NPISHs). Therefore, this definition includes the value of any social contributions and income tax payable on the benefits by the beneficiary to social insurance scheme or tax authorities. To be included in these groups of variables, the transfer must meet two criteria: (1) the coverage is compulsory and (2) it is based on the principle of social solidarity. Importantly, the social benefits included in EU-SILC, with the exception of housing benefits, are restricted to cash benefits.

A. Regional Compensatory and Social-Investment Spending
We now apply the cumulation methodology to obtain-for each one of the selected variable described in the previous section-the NUTS1 level average average of the amount of Euro an household received in a year. We then categorized all these variables into the two groups of compensatory and social-investment variables. The national average over 4 years is reported in Table 4, while in Figure 2, we report the CV indicators computed at European level (EU28) for both total social-investment and total compensatory variables.
First of all, we observe that there is a remarkable difference in the CV for total socialinvestment across Europe, being the CV almost 0.70 in 2007 and much larger in comparison to the CV for total compensatory. However, we also observe that even though the difference for total social-investment remains higher than for total compensatory, there is a tendency for a reduction in the period 2007-2013. In line with Nikolai (2012) and Obinger and Starke (2014), but relying on a very different dataset, we therefore find evidence for a σ − convergence in social-investment spending in Europe, while we  2007,2009,2011,2013) of the amount of euro an household received for each spending category. Data are derived from EU-SILC data through the cumulation methodology (see Section III).

FIGURE 2
European CV-Total Compensatory and Total Social-Investment Spending  Table 5. observe a more stable pattern for total spending for compensatory policy.

V. SVAR ANALYSIS
In this section, we use the dataset described in the previous sections to estimate a SVAR model to identify causal relationships among our variables of interests. SVAR models are among the most prevalent tools in empirical economics to analyze causal effects (Stock and Watson 2007). The underlying setup is the reduced-form vector autoregressive (VAR) model, which is a system of equations for a vector of k variables, in which each variable is made dependent on its own past values, the lagged values of the other variables, and a specific white-noise error term. This model can be easily estimated through standard regression methods (e.g., OLS), since all the regressors are predetermined variables. The reducedform VAR model, however, does not provide enough information to study the causal relationships among the variables and is typically used for the sake of descriptive statistics and forecasting only. It does not provide the structural information because it typically omits the possible influence of contemporaneous values and it delivers error terms that are usually correlated (across variables), so that they cannot be interpreted as genuine shocks affecting the system or as exogenous interventions. Thus, the estimated parameters cannot be used to predict the effect of an intervention. Structural analysis is instead the objective of SVAR models, which add structural information to the VAR (i.e., they solve the identification problem) so that one can recover the causal relationships existing among the variables under investigation. The common approach is to derive this structural information from economic theory or from institutional knowledge related to the data generating mechanism (Stock and Watson 2007).
In the following, we instead rely on a more data-driven approach recently developed in the literature by Moneta et al. (2013) to fully identify the SVAR model. In particular, Moneta et al. (2013) have shown that if the estimated (reduced-form) VAR residuals are non-Guassian, one can exploit higher-order statistics of the data and apply ICA, that is, independent component analysis (Hyvärinen, Karhunen, and Oja 2001). This method has therefore the great advantage of avoiding subjective choices and theory-driven considerations to estimate SVAR model. ICA methods for the statistical identification of SVAR models have also been proposed by Gouriéroux, Monfort, and Renne (2017) and Lanne, Meitz, and Saikkonen (2017). In the following, we briefly review this methodology. For interesting applications of this method, see Brenner et al. (2017); Guerini and Moneta (2017); Ciarli, Coad, and Moneta (2018);and Herwartz (2018).

A. Independent Component Analysis and SVAR Identification
We can denote by Y t = (Y 1t , … , Y kt ) ′ the values at a particular time t of a multiple time series dataset composed of k variables collected for T periods. A simple-but useful-way of representing the data generating process is to model the value of each variable Y kt as a linear combination of the previous values of all the variables as well as their contemporaneous values: where the diagonal elements of the matrix B are set equal to zero by definition and where t represents a vector of error terms with covariance matrix E(ε t ε ′ t ) = ∑ ε . Since these terms represent the structural shocks affecting the system, we can assume that they are uncorrelated, so that ∑ ϵ is a diagonal matrix and that ϵ 1t , … , ϵ kt are mutually independent. Uncorrelatedness of the shocks is a standard assumption in the SVAR literature, while independence is usually not explicitly assumed (also because in a Gaussian setting it is equivalent to uncorrelatedness), but is implicit in many discussions about the economic interpretations of the shocks (Kilian and Lütkepohl 2017).
The model in the standard SVAR form c an be equivalently written as Since variables are endogenous in (3) and (4), this model cannot be directly estimated without biases. It is typical therefore to derive and estimate the VAR reduced form which can be straightforwardly estimated through standard regression methods (e.g., OLS regressions). The problem of identification is therefore the problem of finding the appropriate 0 . Traditionally, this problem is solved by choosing 0 on the basis of a Cholesky factorization of the estimated matrix Σ u of covariance among the reduced-form residuals u t . This imposes a recursive structure among the variables ( 0 results lower triangular) and yields orthogonal structural shocks. A problem with this method, however, is the Cholesky factorization is dependent on the chosen order of the variables (Y 1t , … , Y kt ) ′ in Y t . A reordering of the variable will produce a different Cholesky factorization and a different recursive causal chain among the variables. Thus, this way of proceeding can only be used when the recursive ordering implied by the identification scheme is supported by theoretical or institutional knowledge.
The method proposed by Moneta et al. (2013) instead, applies a search procedure based on ICA, which is able to find, on the basis of data and statistical analysis alone, the appropriate matrix 0 that relates the vector of the structural shocks t such that 0 u t = ϵ t . ICA starts from the consideration that u t are mixtures, that is, linear combinations, of latent sources, or independent components, t . It is crucial for ICA, that t are independent and non-Gaussian. Hence, 0 and ϵ t are recovered by searching the linear combinations of u t that are least statistically dependent in the style of unsupervised statistical learning typical of the machine learning research (Hyvärinen, Karhunen, and Oja 2001), where the measure of statistical dependence used in this context is mutual information. Non-Gaussianity here goes hand-in-hand with independence: if t are non-Gaussian and independent, any linear combination of them will be closer to a Gaussian distribution (see central limit theorem). Then, ICA can also be seen as method which searches for linear combinations of the data that maximizes non-Gaussianity. Hyvärinen, Karhunen, and Oja (2001) show that searching for linear combinations of u t that are maximally independent (or least dependent) is equivalent to searching for t that are maximally non-Gaussian (using the notion of negentropy).
ICA alone, however, leaves undetermined the scale, the sign and order of the latent sources or structural shocks. In other words, −1 0 is identifiable up to a column permutation and the multiplication of each of its diagonal elements by an arbitrary nonzero scalar (Gouriéroux, Monfort, and Renne 2017). While the scale indeterminacy can easily solved by rescaling the column of −1 0 so that all the shocks have unit variance, to solve indeterminacy of the order of the column of −1 0 we need to make further steps, hinging on a further assumption.
Hence, in the following, we rely on a more general identification scheme, called LiNGAM, that is, linear Non-Gaussian acyclic model (Moneta et al. 2013;Shimizu et al. 2006), which incorporates ICA (more specifically, the FastICA algorithm by Hyvärinen, Karhunen, and Oja 2001) in the first step, and then solves its indeterminacy problems by making the further assumption of recursiveness. This assumption means that, given a particular contemporaneous causal order of the variables, the 0 matrix can be transformed in a lower-triangular matrix and the contemporaneous causal order of the variables can be represented as a directed acyclic graph (Moneta et al. 2013). 3 It is important to notice that with LiNGAM, the specific ordering of the variables that produces a lower triangular matrix ( 0 ) is found out directly from the data, while in the Choleski scheme is given a priori. LiNGAM recovers the specific ordering of the variables that produces a lower triangular matrix ( 0 ) from the output of ICA. Since, under recursiveness, both 0 and −1 0 contain k(k − 1)/2 zero entries, LiNGAM search for the unique permutation of −1 0 which has nonzeros on the main diagonal. Since ICA estimates −1 0 with measurement errors, LiNGAM actually searches the permutation which makes −1 0 the closest as possible to lower triangular.
To summarize, our procedure is based on the following assumption: 1. The shocks (ϵ 1t , … , ϵ kt ) are non-normally distributed.
3. The contemporaneous causal structure among (Y 1t , … , Y kt ) is recursive, that is there exists a reordering of the variables such that Γ 0 is lower triangular; the appropriate ordering of the variables, however, is not known to the researcher a priori.
The first assumption can be easily tested in the data. The second assumption is consistent with the interpretation of the elements of ϵ t as structural shocks, that is, exogenous processes that affect each variable of the system at each time in an independent way. In other words, this assumption means that any shock affecting, for example, the level of compensatory spending will not simultaneously affect the shock affecting the level of investment spending (although it can of course also affect the variable level of investment spending). This assumption, however, cannot be directly tested. Finally, the third assumption is necessary to perform the LiNGAM method. While it has the disadvantage of relying on a lower-triangular scheme, LiNGAM has the clear advantage compared to other algorithms of providing a complete identification of Γ 0 (with the entire causal graph of the contemporaneous structure) directly from the data.

B. Results
Relying on NUTS1 level data, we apply the ICA method to explore relationship between the level of compensatory and social-investment spending on the level of NEETs, unemployment and employment of the young. The results from this SVAR analysis are reported in Table 5 and can be interpreted in a causal way. The column variables are the cause, while the row variables are the effects. The model is estimated in differences as variables are highly persistent. To validate the use of this methodology, we conducted checks on the empirical distributions of the VAR residuals (u)-as well as the results of the Shapiro-Wilk and the Jarque-Bera tests for normality; for all the variables, the tests rejects the null hypothesis of normality for the residuals (results are available upon request).
We start by observing the contemporaneous effects from Table 5. It must be noted that the structure of this table reflects the recursive structure implied by the ICA method. After reordering the variables (i.e., NEETs, Employment Young, Log Compensatory, Unemployment Young, Log GDP, and Log Inv), a lower triangular structure emerges. 4 For our purpose, this matrix is not very informative as it implies zero contemporaneous impact of social-investment spending (i.e., Log Inv) on any of our variables of interests, that is, (un-)employment and NEETs, and a significant impact of compensatory spending on GDP.
We therefore resort to an IRF, which describes over a specified time horizon the evolution of the variable of interest after a (one-standard deviation) shock to another variable in the system. In Figure 3, we report the IRFs which are related to our policy variables, that is, the total amount spent in compensatory and socialinvestment policies per household. However, it is important to remark that although we focus our attention on a single shock hitting only one policy variable at time, the shock in the policy variable will consequently affects all our variables in the system. Through the IRF analysis, we thus take into account both the direct and indirect effects (e.g., through a variation in GDP) of a shock in one of our policy variables.
The first thing to notice is that a one shock deviation in the level of compensatory spending per household (about 1,000 Euro) will slightly and significantly increase up to 0.2% the level of NEETs, although this effect tends to become zero and statistically insignificant within three years. On the contrary, a shock in the level of social-investment spending per household (about 1,350 Euro) will slightly reduce the level of NEETs (about −0.2%) although this effect tends to become zero and statistically insignificant over time.
We then observe that the same shock in compensatory spending has no significant effect on employment, while the shock in socialinvestment spending has a small positive and significant effect on it (up to 0.4%). This latter effect tends to disappear after few years. Finally, we observe that the shock in compensatory spending has also a significant and positive effect on unemployment (up to 0.6%), while the shock in investment spending has a significant, although smaller, negative effect on it (up to −1%). We check the robustness of these results by replacing youth unemployment rates with the ratio of youth to adult unemployment rates. Indeed, it has been shown (Pastore 2018) that this indicator is less affected by the fluctuations of the economic cycle and more closely reflect a country's institutional characteristics. While no significant effects emerge for the relative unemployment rates, results (available upon request) confirm that investment spending has a positive effect on NEETs and employment rates.
Overall, these results suggest that shocks in the level of total investment spending lead to positive economic outcomes. Two remarks, however, are in order. First of all, as highlighted above, through the IRFs we are observing both the direct and indirect effects of a shock in a policy variable. That is, our results suggest that a shock in the level of compensatory spending does not ultimately lead (e.g., through an increase in GDP) to a significant reduction in the level of NEETs and unemployment. This result is in line with the analysis of Bruno, Marelli, and Signorelli (2015) who found that both NEETs and unemployment rates respond slowly to an increase in GDP, with many years elapsing before the situation of the young improves. In addition, as recent studies on intrafamilial transfers suggest (Gál, Vanhuysse, andVargha 2018, Francesconi andHeckman 2016), there is a danger in looking at data on public transfer alone without considering intergenerational transfers (cash) and the household economy (time), as social investments may have a differential impact across childhood and early youth through family investments.

VI. CONCLUSIONS
As it has been already highlighted, both in the literature and at the institutional level, the regional dimension does matter. There are strong differences across regions in EU, but also inside individual countries. Therefore, in order to better target policy measures, there is an increasing need of social policy indicators developed at regional level. Moreover, since young people paid the highest price during the global economic crises, there is also a renewed sense of urgency to integrate them into the labor market and into the education system. Our paper offers contributions in both respects: we construct new indicators of regional spending and we investigate their impact on new indicators-such as NEETs-of youth labor market participation.
In particular, we relied on SAE techniques, as applied to the EU-SILC survey, to develop new indicators of compensatory and socialinvestment spending at NUTS1 level. These methodologies have already been proved to be successful to derive regional measures of poverty (Betti et al. 2012;Verma, Gagliardi, and Ferretti 2013). Interestingly, by looking at these measures, we can observe across EU Member States regional convergence of compensating expenditure, and a milder of social investment.
We then used these new regional indicators of spending in combination with a recently developed SVAR approach (Moneta et al. 2013;Shimizu et al. 2006) to investigate the causal relationships between labor market outcomes and different types of spending. While relying on independent component analysis, this method has the great advantage of avoiding subjective choices and theory-driven considerations to estimate SVAR model (Gouriéroux, Monfort, and Renne 2017;Lanne, Meitz, and Saikkonen 2017). Our main result suggests that social-investment policies strongly differ across EU regions but can be effective to enhance labor market outcomes of the young. Indeed, it is possible to observe an improvement in both NEETs and employment rates after a shock in social investment. Moreover, in line with the analysis of Bruno, Marelli, and Signorelli (2015), our results suggest that youth labor market indicators are less responsive to variation in compensatory spending as the total effects of a shock in these policy variables do not ultimately lead to an increased participation of the youth in the labor market. These results also highlight the importance of explicitly considering the role of intrafamilial transfers in future analysis as social investments may have a differential impact across childhood and early youth through family investments (Gál, Vanhuysse, andVargha 2018, Francesconi andHeckman 2016). At the same time, we need to be cautious as we did not consider any redistributive/differential effect that such policies can have at the individual levels (Bonoli and Liechti 2018;Pavolini and Van Lancker 2018), nor any regional spillover effects that policies can have among regions. Future research thus need to further investigate possible complementaries between the two types of policies as well as the explicit role of family investments (both in cash and in time), as youth employment remains the crucial node to sustainable economic and social development.

APPENDIX: A EU-SILC VARIABLE SELECTION
Variable included in compensatory spending: 1. Unemployment benefits (PY090G): refer to (full o partial) benefits for benefits compensating for loss of earnings. It also includes early retirement, vocational training, redundancy compensation, severance, and termination payments; 2. Old-age and survivors benefits (PY100G): refer to the provision of social protection against the risk linked to old age