Childhood-Related Policies and Adult Poverty: Evidence from Some European Countries

For an overview of EU-SILC, see Wolff et al. (2010). To access further information about EU's regulations concerning the SILC, data documentation provided by Eurostat, and SILC variable lists, we recommend the EU-SILC web portal provided by the GESIS research institute at http://​www.​gesis.​org/​.

The Comparative Welfare Entitlements Dataset (CWED) provides systematic data on institutional features of social insurance programs in 33 countries spanning much of the post-war period. The original data sources of most variables are the OECD's Social Expenditure Database and the International Labor Organization (ILO). For an overview of the CWED, variable lists, data documentation, original data sources and list of publications using this dataset see http://​cwed2.​org/​.

Bryan and Jenkins (2013) review the two-step approach and other modeling techniques recently appeared to deal with multi-country dataset, that is, individual units sampled within groups or clusters (countries).

As measures of social exclusion, material deprivation and material hardship are becoming more and more available, the concept of multidimensional poverty is increasingly used. For example, Figari (2012) analyses the relationship between deprivation, income and other individual dimensions over time, in eleven European countries. He finds that a relevant part of the deprivation gap is attributable to a country specific effect revealing the importance of unobserved factors like institutions, which is the focus of this paper. Visser et al (2014) study to what extent macro-economic circumstances and social protection expenditure affect economic deprivation in several European countries. They find that the strength of the relationship between a low income and economic deprivation strongly varies according to the economic circumstances in a country and the generosity of the welfare state. For an overview of this literature see Nolan and Marx (2009) or Morelli et al (2015) and references therein.

Corak (2006) provides a careful and interesting discussion on the rationale of these two public policy devices and the underlying assumptions behind their mechanisms.

Public spending in other social policy areas such as health care and housing could also be considered. However, to the extent that they have a wider set of beneficiaries, in particular the elderly in the case of health spending and even they might not exclusively assist families, it could be hazardous to isolate its long-run impact on children which is the main objective of the paper. See for instance, Barcena-Martin et al. (2017) for evidence on the impact of other social transfer on child poverty reduction.

The relative effectiveness of targeted versus universal programs have also been intensely debated. Corak et al (2005) find that the best-performing countries in child poverty reduction tend to have systems of universal child benefits and tax concessions that are not particularly strongly targeted at low-income children. Although it is certainly important, this debate falls beyond the scope of this paper. See Marx et al. (2014) for a survey of studies on child benefits. See also Corak (2013) for a detailed review on the empirical evidence for the impact of family allowances on poverty transmission.

Sylwester (2000) in a cross-country analysis finds that higher levels of initial income inequality are associated with higher public education expenditure. See also Corak (2006) or Marx et al. (2014) for additional discussion on the relationship between inequality and social policies.

In contrast to the US, where the poverty measure is an absolute standard, in Europe the most common measure is a relative one (see Burkhauser 2009 for a discussion on this issue). Indeed the official poverty measure in the European Union is the one used in the paper, see http://​ec.​europa.​eu/​eurostat/​statistics-explained/​index.​php/​Glossary:​At-risk-of-poverty_​rate. Alternative measures of monetary poverty could be considered. For instance, most current social transfers could be excluded from the income definition here. Results using that measure, available upon request, are in line with the findings in the paper.

The intergenerational modules do not consider as eligible for inclusion in the survey individuals who are not the selected respondent in register countries and individuals not in the age range (25–65). For additional details see http://​ec.​europa.​eu/​eurostat/​web/​income-and-living-conditions/​methodology/​list-variables.

This information includes family composition, year of birth of parents, occupation and level of education of parents. We do not use information on parents’ occupation, since these variables contain a large fraction of missing values. See http://​ec.​europa.​eu/​eurostat/​web/​income-and-living-conditions/​data/​ad-hoc-modules.

To assess the long-run effect of household characteristics, we exclude those individuals who lived in a collective house or in some institution when young.

See http://​www.​uis.​unesco.​org/​Education/​Pages/​default.​aspx and http://​data.​worldbank.​org/​ for data on public expenditure in education and PPP, respectively.

Since our sample contains 13 countries, in principle we could have data for corresponding to 13 ×  30 = 390 distinct country-year cells. However, we lack data individual data corresponding to 2 of these cells. This can be due to the fact that EU-SILC did not collect data for all individuals in our sample. In addition, we lack aggregate level data for several country-cohorts. For instance, regarding public expenditure in primary education we lack data corresponding to 95 cells. This is due to the fact that UNESCO did not collect data every year in every country. We use a simple linear interpolation to smooth expenditure data. We cannot do this when missing data for a given country correspond to the first or the last years of the period considered. As an example, data for Belgium are only from 1975 onwards while for Spain they are from 1972. Therefore, by interpolating, we recover all but 20 cells for public expenditure in primary education and 169 cells for family allowances.

Alternative datasets, as the available updates of the ILO and the EUROSTAT were considered but dismissed for several reasons. Unfortunately, the updates of the ILO-The Cost of Social Security for the period after 1990 are not appropriate as they revised both the methodology and framework from previous inquires which prevents from having comparable figures. EUROSTAT database is only available from 1990 onwards, and combining it with the CWED is not possible as the definition of “family allowances” does not exactly coincide in both sources (see ILO 1997; ILO 2017).

See Grogger (1996), Mayer and Lopoo (2008) and, more recently, Hidalgo-Hidalgo and Iturbe-Ormaetxe (2018).

With our measures, for those born in 1956 we assign public spending in 1971 only (they are 15 in 1971), while for those born in 1957 we assign the average of public spending in 1971 and 1972 (they are 14 and 15, respectively). In addition observe that the availability of data on aggregate variables implies that the age range is not the same for both cross-sections (25–49 and 25–55 in 2005 and 2011, respectively). This might introduce some bias in the results. We explored the possibility of considering the same age range in both waves (see Sect. 5 below).

We are aware that not considering immigrants might provide an incomplete view of current poverty situation in Europe. Unfortunately the EU-SILC database does not provide information about the year of migration without which we cannot claim whether or not these individuals were exposed to public policies.

See also Donald and Lang (2007) for a detailed discussion of the statistical properties of the two-step estimator (compared to Generalized Least Squares).

Different econometric methods can be used instead, for instance, multilevel approach. To check the robustness of our results we also used a random-intercept model in which the intercept is allowed to vary between country-cohorts. The results obtained (not shown to keep space at a minimum but available upon request) are consistent with the findings discussed in the main analysis below.

Nevertheless we replicate our results by reducing the size of the clusters, in particular we define 5-years country-cohorts. This implies reducing the number of groups to 78 (13 countries and 6 five-year periods) but increasing their size to around 1873 individuals. Results, available upon request, are very much in line with the ones found with our group definition.

We could alternatively consider expenditure on education and family allowances as a share of GDP (see Hidalgo-Hidalgo and Iturbe-Ormaetxe 2018 for a comment of this approach). However, we choose our measure for two reasons. First, it is a more general specification (observe that using the ratio implies restricting the effect of GDP to be the mirror effect of the expenditure). Second, it provides a more clear interpretation of results (since a large value in this ratio can be due either to high spending or to low per head GDP).

Note that, as we have two cross sections, these regressors take different values for each combination of year of birth, country and cross section. See the “Appendix” for a detailed description of all the variables we use and the bottom part of Table 1 for descriptive statistics of these aggregate variables.

Instead, we could jointly estimate both public policies (expenditure in education and in family allowances) to see net effects. Nevertheless, this approach has two important shortcomings. Observe that proceeding in that manner drastically reduces the number of country-cohorts. The reason is that data on both public policies, and in particular family allowances, is not available for all country-cohorts we have in the sample: we lack data for all cohorts from Spain and Greece and cohorts born in and after 1977 for all countries. This means that results are not directly comparable to those in the rest of the paper. Moreover, unfortunately, both spending categories might be highly correlated which makes difficult to estimate their effects separately (the coefficient correlation is equal to 0.17 and it is statistically significant at the 1% level). Indeed, Mayer and Lopoo (2008) also acknowledge this problem and estimate the main model for the different spending categories separately.

See Nolan and Marx (2009) for similar evidence on the impact of the number of dependant on poverty status and Valletta (2006) for poverty incidence in married versus single individuals.

Parental education is measured with a dummy equal to one when at least one of the parents has secondary education. Requiring tertiary education would be too restrictive as only 12.68% of individuals in the sample have at least one parent with tertiary education. We also explored the possibility of introducing separately the educational levels of both parents. The results are very similar, although the sample size gets much lower because of the large increase in missing values: the dummy variable capturing parental education is missing only when the education level of both the father and the mother is missing. See “Appendix C” in the Online Appendix.

Public spending policies might have a different impact on poverty reduction depending on the country GDP and or income inequality. In “Appendix E” we analyze this possibility and find that it is not the case here.

We perform a robustness check to analyze whether there is one specific country driving the main findings in the paper. Results, shown in “Appendix F”, confirm that by excluding one country each time we get findings very close to those in Table 4.

This is the difference between \( {{\Phi }}\left( {\overline{\text{X}} \widetilde{\text{B}}} \right) \) and \( {{\Phi }}\left( {\overline{\text{X}} \widehat{\text{B}}} \right) \) where \( \overline{\text{X}} \) denotes the vector of means of all controls in Eq. (1), \( \widetilde{\text{B}} \) is the vector of estimated coefficients in Eq. (1) and \( \widetilde{\text{B}} \) is equal to \( \widehat{\text{B}} \) except for the estimated coefficients of the country-cohort dummies, \( \widetilde{\beta }_{ct} , \) in particular, \( \widetilde{\beta }_{ct} = \widehat{\beta }_{ct} - 0.0963 \).

Recent research suggests the need to consider institutional indicators as replacement rates or measures of pensions as additional measures of policy effort and design. See Marx et al. (2014) for a comprehensive analysis on institutional indicators versus spending-based measures.

Evidence on the relationship between current labor market performance and poverty rates is mixed. At the aggregate level, employment rates are not the main driver of cross-country differences in the overall poverty risk among the working aged population (see Marx et al. 2014; OECD 2009).

The experimental Perry Preschool Program in the US in the 1960s provided schooling and home visits to children aged 3–4 year from poor families. The Abecedarian Project recruited children with high risk of social exclusion age 4.4 months on average who attended a preschool center. Head Start was a similar national pre-school program, which targeted a package of education, health and nutrition services to poor families.