## Executive Summary

## Source details

**Canonical URL:** [Executive Summary](https://www.imf.org/-/media/files/publications/wp/2023/english/wpiea2023134-print-pdf.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/wp/2023/english/wpiea2023134-print-pdf.pdf.md)
- [Structured JSON version](/-/media/files/publications/wp/2023/english/wpiea2023134-print-pdf.pdf.json)

---

### Key Findings
- During the COVID-19 pandemic, child poverty increased dramatically in the European Union.
- In 2020 alone, the number of children suffering from severe material deprivation increased by 19 percent (0.9 million) in the EU.
- The increase in child poverty in 2020 was 3¼ times more than the increase in adult poverty.
- Large cross-country heterogeneity:
  - About half of EU countries experienced an increase in child poverty in 2020 while the other half experienced a decline.
  - The sharp EU-level increase was driven by dramatic rises in countries with large child populations; for example, the number of children affected by severe material deprivation almost tripled in Germany in 2020.
  - The share of children suffering from poverty increased by 3.0 to 4.0 percentage points in Germany, Romania, and Spain; together these three countries account for almost a third of EU children.
- Child poverty declined between 2012 and 2019 (the share suffering from severe material deprivation was cut by half between 2012 and 2019), but the pandemic reversed this trend.
- The educational loss from school and childcare closures during the pandemic was more severe for the poorest children, compounding scarring risks.

### Long-term and Macroeconomic Implications
- Child poverty has persistent negative consequences for individuals: impairing skills development, health, educational achievement, and future earnings and well‑being.
- Macro-level costs include higher remediation costs, lower human capital accumulation, adverse effects on productivity and potential growth, reduced employability, higher inequality, lower social mobility, and fiscal costs.
- Evidence on aggregate cost estimates (literature examples):
  - United States: estimated economic cost of child poverty reaches "4 percent of GDP" (Holzer and others (2008)).
  - McLaughlin and Rank (2018) estimate the cost of child poverty at "5.4 percent of GDP."
  - Reduced productivity accounts for about "30 percent" of the estimated total cost of child poverty.
- The duration of poverty matters: long-lasting child poverty amplifies scarring and is costlier to remediate later; earlier poverty reduction is less costly than later remediation.
- Early childhood investments yield high returns: randomized-controlled experiments show spending on birth-to-five quality education programs for disadvantaged children produce a "7–13% annual return on investment."

### Drivers and Channels of Rising Child Poverty
- Cross-country differences in child poverty dynamics since the GFC are strongly associated with:
  - Structural features: income inequality (Gini), average household size, share of children with single parenthood.
  - Economic cycle effects, notably via labor market impacts (unemployment, output gap, low work intensity).
  - The generosity and design of social protection spending on family and children.
- The pandemic’s impact varied with the severity of the economic shock, differences in policy responses, existing safety nets, and structural characteristics.
- COVID-19 specific variables: the "stringency of containment measures" performed best among pandemic indicators and was significant for younger children, indicating childcare/school closures affected parents’ capacity to work.

### Labor Policy Options
- Policies to increase parental employment and labor income can reduce child poverty and mitigate scarring:
  - Reduce obstacles to work: increase working hours flexibility, promote work-life balance, reduce gender biases in employment, and increase access to childcare for low-income parents.
  - Facilitate combining work and parental responsibilities to increase working hours and labor force participation—particularly impactful for single parents.
  - Adjust tax-benefit design to increase financial incentives for parents to work and boost women’s labour force participation.
  - Increase net wages of poor parents—for example:
    - Reduce payroll taxes paid by employees at the low end of the pay scale to increase net wages without raising employers’ labor costs.
    - Reduce payroll taxes paid by employers at the low end of the pay scale to increase job opportunities without reducing net wages.
  - Consider tax-credit instruments similar to the U.S. Earned Income Tax Credit—evidence shows increases in labor supply (notably single parents), reductions in poverty, and improvements in children’s health and human capital.
- Stable and predictable income reduces child poverty levels and mitigates long‑term impacts by reducing mental stress, improving parenting, improving infant brain activity (randomized evidence), and increasing school attendance.

### Fiscal Policy Options and Design
- Social protection spending on family and children can have a powerful and immediate impact on child poverty.
- Constraints: fiscal tightening in many countries to reduce deficits and public debt and to fight inflation makes increasing spending challenging.
- Repeating post-GFC consolidations that reduced social protection spending on family and children would delay poverty reduction and increase scarring.
- Increased means-testing implemented post-GFC did not appear to increase the impact of spending on child poverty and is unlikely to do so post-COVID-19 unless design and implementation issues are addressed.
- Short-term priorities:
  - Preserve or, when possible, increase spending on family and children.
  - Initiate reforms to increase the impact of such spending on child poverty, including:
    - (i) introducing a universal child benefit taxed at a sufficiently progressive and broad-based personal income tax,
    - (ii) making cash transfers more conditional on caregivers taking action for the child’s wellbeing (e.g., school attendance, health check),
    - (iii) increasing the amount of transfers to specific groups more affected by child poverty such as single parents or parents of younger children.
- Fiscal evidence (regression highlights): family and children spending (share of GDP) consistently shows a negative association with severe material deprivation (examples of coefficients: -0.309**** (0.061); streamlined specification -0.410**** (0.095)).

### Public Investment and Complementary Policies
- Policies to mitigate long-term impacts require public investment beyond income support:
  - Invest to increase availability of childcare, notably in poor neighborhoods, complementing cost-reduction initiatives.
  - Increase investment in education to improve inclusion of poor children and alleviate pandemic-related learning loss; education spending at EU level declined by 0.4 percentage points of GDP to 4.7 percent of GDP between 2010 and 2019; capital spending on education averaged 0.4 percent of GDP during the period.
  - Improve parents’ skills to reinforce labor policy effects on employability and labor income.
  - Increase access to healthcare and medicine for children and pregnant women.
  - Provide adequate nutrition to poor children and pregnant women (FEAD is primarily a food aid scheme with 80 percent of its budget devoted to food).
  - Implement housing policies to address contextual determinants of poverty while avoiding concentration of disadvantage.
- Complementarity: cash transfers and in-kind benefits are complementary; in-kind spending is more efficient for younger children in some specifications, but non-means-tested spending generally has larger impacts than means-tested spending.

### Policy Priorities and Sequencing
- Dual objectives: (1) rapidly reduce the level of child poverty; (2) mitigate its long-term impact.
- Policies should be broad-based and multidimensional—income support alone is insufficient given multiple interrelated channels.
- Labor and fiscal measures should be implemented as a coherent package, recognizing complementarities and trade-offs under constrained fiscal space.
- Immediate actions: preserve/increase family-and-children social protection spending and target reforms to improve effectiveness.
- Medium- and long-term actions: structural labor market reforms, public investment in childcare/education/health, skills programs, and housing policies to prevent scarring.

---

### Empirical Evidence and Key Statistics

### Aggregate and country-level trends
- The share of children suffering from severe material deprivation increased markedly in 2020, "erasing, in one year, all the reduction achieved in the second half of the 2010s."
- Partial evidence suggests the number of children suffering from poverty is likely to have further increased in 2021; at time of writing, severe material deprivation data for 2021 were not available, but "the at-risk-of-poverty rate rose by 0.3 ppt on top of a 0.7 ppt increase in 2020."
- "Child poverty may have further increased in 2022 as more people may be unable to afford essential goods due to high inflation (notably of food and energy)."
- The number of children suffering from severe material deprivation rose in 12 EU countries in 2020 but decreased in 14 others.
- Between 2019 and 2021, the number of children at risk of poverty increased in 13 EU countries but declined in 14 others.
- Severe material deprivation across EU27 ranges from "3.3 percent in Slovenia and Greece to 29.3 percent in Estonia."
- Selected country figures (Number / In percent of EU total / In 2020 (percent) / Change 2019-20 (ppt):
  - Germany: "13,677,902 16.8 5.9 3.8"
  - Romania: "3,644,619 4.5 21.4 3.7"
  - Spain: "8,325,756 10.2 9.0 3.0"
  - France: "14,540,168 17.9 6.3 1.2"
  - Italy: "9,433,159 11.6 5.7 -0.8"
  - Poland: "6,913,237 8.5 2.3 -0.3"
  - Sweden: "2,155,379 2.6 3.0 -0.1"
  - Bulgaria: "1,189,680 1.5 20.0 -0.2"
  - Belgium: "2,320,244 2.9 4.6 -0.8"
  - Netherlands: "3,337,245 4.1 2.7 0.3"
  - Aggregate totals (Table 2 groupings):
    - Total (Severe Material Deprivation increased in 2020 group): "50,322,380 61.8 7.8 2.3"
    - Total (Severe Material Deprivation decreased in 2020 group): "30,901,937 38.0 5.1 -0.8"
    - Luxembourg (Severe Material Deprivation remained unchanged): "119,539 0.1 1.3 0"
- In 2020, the number of children suffering from severe material deprivation increased by 19 percent (0.9 million) in the EU.

### Differences between poverty measures and age patterns
- Severe material deprivation and at-risk-of-poverty provide different pictures. Example: Romania experienced an increase in severe material deprivation (+ 3.7 ppt in 2020) while the share of children at risk of poverty declined (- 0.7 ppt in 2020 and - 2.0 ppt if anchored at 2008).
- Age-differentiated changes (EU27, in percentage points) — Severe material Deprivation (2019-20): "1.1 1.1 1.0 1.2" (headers correspond to age groups: "Less than 18 Less than 6 6 to 11 12-17").
- Table excerpts: At-risk-of-poverty rate (2019-21) "1.0 1.9 0.9 0.0".

### Panel regressions and machine learning highlights
- Period covered: 2009–2020; cutoff date for data: November 22, 2022.
- Estimation approach: pooling, fixed effects, and random effects tested; pooling and Hausman tests led to selection of the fixed effects model; heteroscedasticity-consistent covariance estimators used.
- Key regression associations (selected coefficients preserved):
  - Stringency of containment during pandemic: 0.161**** (0.043) for All Children (Baseline (1)); 0.203**** (0.051) for Younger Children (Baseline (1)).
  - GFC dummy examples: 0.188**** (0.054); 0.247**** (0.063).
  - Gini coefficient: 0.101* (0.055); 0.126** (0.060).
  - Average household size: 0.354**** (0.058); 0.379**** (0.076); 0.356**** (0.059).
  - Share of children with single parent: 0.203** (0.080); 0.220*** (0.072); 0.285*** (0.108).
  - Unemployment rate: 0.229**** (0.043); 0.237**** (0.046); 0.236**** (0.048).
  - Share of children in households with very low work intensity: 0.174*** (0.062); 0.194*** (0.066); 0.160**** (0.082); 0.178**** (0.050).
  - Temporary employment share: 0.173* (0.090); 0.211** (0.102); 0.268* (0.139).
  - Family and children spending (share of GDP): -0.309**** (0.061); additional streamlined/specification examples: -0.410**** (0.095); -0.425**** (0.117).
  - Observations and fit: examples include Observations: 322 and 310; R-Squared: 0.623, 0.627, 0.533, 0.539; Adjusted R-Squared: 0.582, 0.579, 0.478, 0.481.
- Streamlined specification (Table II.2) — Severe Material Deprivation:
  - All Children (Less than 18): Output gap -0.085**** (0.043); Share of children with single parent 0.377*** (0.130); Family and children spending -0.410**** (0.095); Observations: 320; R-Squared: 0.552.
  - Younger Children (Less than 6): Output gap -0.112**** (0.028); Share of children with single parent 0.413**** (0.076); Share of children in households with very low work intensity 0.270*** (0.049); Family and children spending -0.425**** (0.117); Observations: 320; R-Squared: 0.439.
- Machine learning models (comparison, Table II.4) show strong predictive performance (selected results):
  - All children — R2 examples: Linear regression 0.919; Elastic net 0.926; K-nearest neighbors 0.945; Support vector regression 0.939; Random forest 0.937; Extreme Gradient Boosting 0.934.
  - Small children — R2 examples: Linear regression 0.863; Elastic net 0.879; K-nearest neighbors 0.900; Random forest 0.902.
- ML main determinants averaged across models: household characteristics and income, labor market participation, and family and children benefits; additional important determinants: low childcare enrollment and low parental education.
- ML threshold findings (random forest): child poverty declines until GDP per capita reaches 28,000 PPS; significant reduction of child poverty when family and children benefits reach around 3% of GDP.

---

### Policy Recommendations (Concise and Prioritized)

- Immediate / Short term
  - Preserve current level of social protection spending on family and children; where fiscal space allows, prioritize increases targeted to reduce child poverty rapidly.
  - Deploy fiscal transfers and family-targeted spending to achieve immediate reductions in severe material deprivation.
  - Protect and expand in-kind services (childcare/ECEC) especially in poor neighborhoods to support parental employment and child development.

- Medium term
  - Implement labor market reforms to increase parental employment, working hours, and job stability: reduce barriers to work (flexible hours, leave policies), expand affordable childcare, and improve incentives via tax-benefit design.
  - Consider targeted net-wage support for low-paid parents (reduce employee payroll taxes at low end; consider employer payroll tax reductions to stimulate hiring of low‑skilled workers).
  - Expand parental skills and active labor market programs to improve employability and earnings potential.

- Longer term / Design reforms
  - Consider introducing a universal child benefit taxed through a sufficiently progressive broad-based personal income tax to concentrate net benefits on low-income families while minimizing stigma and complexity.
  - If not universal, redesign means-testing and conditionality carefully to avoid low take-up, perverse withdrawal effects, and work disincentives; consider increasing transfers to high-need groups (single parents, parents of younger children).
  - Scale public investment in education (including recouping pandemic learning loss), health and nutrition for children and pregnant women, childcare infrastructure, and affordable housing to limit scarring and intergenerational transmission of poverty.

---

*Italic: Executive Summary — IMF Working Paper (wpiea2023134-print-pdf)*

### Executive Summary ......................................................................................................

### Executive Summary

### Key Findings
- During the COVID-19 pandemic, child poverty increased dramatically in the European Union.
- In 2020 alone, the number of children suffering from severe material deprivation increased by 19 percent (0.9 million) in the EU.
- The increase in child poverty in 2020 was 3¼ times more than the increase in adult poverty.
- Large cross-country heterogeneity:
  - About half of EU countries experienced an increase in child poverty in 2020 while the other half experienced a decline.
  - The sharp EU-level increase was driven by dramatic rises in countries with large child populations; for example, the number of children affected by severe material deprivation almost tripled in Germany in 2020.
  - The share of children suffering from poverty increased by 3.0 to 4.0 percentage points in Germany, Romania, and Spain; together these three countries account for almost a third of EU children.
- Child poverty declined between 2012 and 2019 (the share suffering from severe material deprivation was cut by half between 2012 and 2019), but the pandemic reversed this trend.
- The educational loss from school and childcare closures during the pandemic was more severe for the poorest children, compounding scarring risks.

### Long-term and Macroeconomic Implications
- Child poverty has persistent negative consequences for individuals: impairing skills development, health, educational achievement, and future earnings and well-being.
- Macro-level costs include higher remediation costs, lower human capital accumulation, adverse effects on productivity and potential growth, reduced employability, higher inequality, lower social mobility, and fiscal costs.
- The duration of poverty matters: long-lasting child poverty amplifies scarring and is costlier to remediate later; earlier poverty reduction is less costly than later remediation.

### Drivers and Channels of Rising Child Poverty
- Cross-country differences in child poverty dynamics since the GFC are strongly associated with:
  - Structural features (income inequality, average household size, share of children with single parenthood).
  - Economic cycle effects, notably via labor market impacts.
  - The generosity and design of social protection spending on family and children.
- The pandemic’s impact varied with the severity of the economic shock, differences in policy responses, existing safety nets, and structural characteristics.

### Labor Policy Options
- Policies to increase parental employment and labor income can reduce child poverty and mitigate scarring:
  - Reduce obstacles to work: increase working hours flexibility, promote work-life balance, reduce gender biases in employment, and increase access to childcare for low-income parents.
  - Facilitate combining work and parental responsibilities to increase working hours and labor force participation—particularly impactful for single parents.
  - Adjust tax-benefit design to increase financial incentives for parents to work and boost women’s labour force participation.
  - Increase net wages of poor parents—for example:
    - Reduce payroll taxes paid by employees at the low end of the pay scale to increase net wages without raising employers’ labor costs.
    - Reduce payroll taxes paid by employers at the low end of the pay scale to increase job opportunities without reducing net wages.

### Fiscal Policy Options and Design
- Social protection spending on family and children can have a powerful and immediate impact on child poverty.
- Constraints: fiscal tightening in many countries to reduce deficits and public debt and to fight inflation makes increasing spending challenging.
- Repeating post-GFC consolidations that reduced social protection spending on family and children would delay poverty reduction and increase scarring.
- Increased means-testing implemented post-GFC did not appear to increase the impact of spending on child poverty and is unlikely to do so post-COVID-19 unless design and implementation issues are addressed.
- Short-term priorities:
  - Preserve or, when possible, increase spending on family and children.
  - Initiate reforms to increase the impact of such spending on child poverty, including:
    - (i) introducing a universal child benefit taxed at a sufficiently progressive and broad-based personal income tax,
    - (ii) making cash transfers more conditional on caregivers taking action for the child’s wellbeing (e.g., school attendance, health check),
    - (iii) increasing the amount of transfers to specific groups more affected by child poverty such as single parents or parents of younger children.

### Public Investment and Complementary Policies
- Policies to mitigate long-term impacts require public investment beyond income support:
  - Invest to increase availability of childcare, notably in poor neighborhoods, complementing cost-reduction initiatives.
  - Increase investment in education to improve inclusion of poor children and alleviate pandemic-related learning loss.
  - Improve parents’ skills to reinforce labor policy effects on employability and labor income.
  - Increase access to healthcare and medicine for children to magnify benefits of education and labor policies.
  - Provide adequate nutrition to poor children and pregnant women to rapidly improve health and mitigate long-term effects.
  - Implement housing policies to address contextual determinants of poverty (where and how people live).

### Policy Priorities and Sequencing
- Dual objectives: (1) rapidly reduce the level of child poverty; (2) mitigate its long-term impact.
- Policies should be broad-based and multidimensional—income support alone is insufficient given multiple interrelated channels.
- Labor and fiscal measures should be implemented as a coherent package, recognizing complementarities and trade-offs under constrained fiscal space.
- Immediate actions: preserve/increase family-and-children social protection spending and target reforms to improve effectiveness.
- Medium- and long-term actions: structural labor market reforms, public investment in childcare/education/health, skills programs, and housing policies to prevent scarring.

_Italic: Executive Summary — IMF Working Paper (wpiea2023134-print-pdf)_

### 6.8 percent. As children fell more into

### 6.8 percent. As children fell more into

### Pandemic impact on child poverty: aggregate and recent trends
- The share of children suffering from severe material deprivation increased markedly, "erasing, in one year, all the reduction achieved in the second half of the 2010s" (Figure 2).
- Partial evidence suggests the number of children suffering from poverty is likely to have further increased in 2021.
- At time of writing, data for severe material deprivation were not available for 2021, but "the at-risk-of-poverty rate rose by 0.3 ppt on top of a 0.7 ppt increase in 2020 (Figure 2)."
- "Child poverty may have further increased in 2022 as more people may be unable to afford essential goods due to high inflation (notably of food and energy)."

### Heterogeneity across EU countries
- "The number of children suffering from severe material deprivation rose in 12 EU countries in 2020 but decreased in 14 others."
- "Between 2019 and 2021, the number of children at risk of poverty increased in 13 EU countries but declined in 14 others."
- The increase in the EU27 as a whole is explained by a sharp increase in child poverty in most large EU members.
- "The share of children suffering from severe material deprivation increased by 3 percentage points or more in a year in Germany, Romania, and Spain (which together account for almost one-third of EU child population)."
- "In Germany alone, the number of children suffering from severe material deprivation almost tripled."

### Country-level figures (selected excerpts from tables)
- Aggregate totals (from Table 2 groupings):
  - Total (Severe Material Deprivation increased in 2020 group): "50,322,380 61.8 7.8 2.3"
  - Total (Severe Material Deprivation decreased in 2020 group): "30,901,937 38.0 5.1 -0.8"
  - Luxembourg (Severe Material Deprivation remained unchanged): "119,539 0.1 1.3 0"
- Selected country rows (Number / In percent of EU total / In 2020 (percent) / Change 2019-20 (ppt)):
  - Germany: "13,677,902 16.8 5.9 3.8"
  - Romania: "3,644,619 4.5 21.4 3.7"
  - Spain: "8,325,756 10.2 9.0 3.0"
  - France: "14,540,168 17.9 6.3 1.2"
  - Italy: "9,433,159 11.6 5.7 -0.8"
  - Poland: "6,913,237 8.5 2.3 -0.3"
  - Sweden: "2,155,379 2.6 3.0 -0.1"
  - Bulgaria: "1,189,680 1.5 20.0 -0.2"
  - Belgium: "2,320,244 2.9 4.6 -0.8"
  - Netherlands: "3,337,245 4.1 2.7 0.3"
  - (Full country-level entries are presented in the source tables.)

### Differences between poverty measures
- "Severe material deprivation and at risk of poverty provide a different picture of the pandemic impact of child poverty."
- Example: "Romania experienced one of the largest increases in the share of children suffering from severe material deprivation (+ 3.7 ppt in 2020) but the share of children at risk of poverty declined (- 0.7 ppt in 2020 and - 2.0 ppt if the poverty line is anchored at its 2008 level)."
- Age-differentiated changes (EU27, in percentage points) — Table 3:
  - Severe material Deprivation (2019-20): "1.1 1.1 1.0 1.2"
  - At Risk of Poverty (2019-21): "1.0 1.9 0.9 0.0"
  - (Headers in Table 3 correspond to age groups: "Less than 18 Less than 6 6 to 11 12-17".)

### Rationale for using severe material deprivation as primary measure
- Severe material deprivation is preferred for cross-country analysis because it is "not sensitive to differences in income levels across EU countries."
- Severe material deprivation measures the "enforced inability (rather than the choice not to do so) to afford some goods considered as essential."
- The at-risk-of-poverty measure is income-relative and sensitive to median income differences and time variation; it complements material deprivation to provide additional insights.

### Long-term impact of child poverty: individual and macroeconomic channels
- Individual-level impacts documented in the literature:
  - Poverty is associated with lower educational achievements, worse physical and mental health, and behavioral issues.
  - "Children in poor families have [...] been documented to have reduced volumes in the cerebral cortex and hippocampus" affecting executive function, language, and memory.
  - Early language gaps emerge by 15–16 months; by age three, "the vocabulary of a child living in poverty is less than half the size of vocabulary of a child from a professional family."
  - Poverty-related stress, material hardship, and reduced access to healthcare contribute to health consequences and educational setbacks.
- COVID-19 specific aggravators:
  - School and nursery closures reduced children's accumulation of human capital; educational loss was "more severe for the poorest children."
  - The pandemic’s impact on skill acquisition makes long-term effects "likely to be larger than during a typical economic crisis."
- Macroeconomic costs:
  - Evidence for the United States: Holzer and others (2008) estimate economic cost of child poverty reaches "4 percent of GDP" due to reduced productivity, crime, and health expenditure.
  - McLaughlin and Rank (2018) estimate the cost of child poverty at "5.4 percent of GDP."
  - Reduced productivity accounts for about "30 percent" of the estimated total cost of child poverty.
- Scarring channel:
  - Rise in child poverty contributes to "scarring—defined as diminished longer-term output relative to pre-pandemic projections" via reduced future productivity and amplified human capital losses from school/nursery closures.

### Policy implications and priorities
- Overarching priorities:
  1. "Reverse as soon as possible the increase in child poverty experienced during the pandemic" because long-term consequences increase with duration in poverty. Severe material deprivation is likely to persist beyond the pandemic, especially with high inflation in 2022–23.
  2. Complement poverty reduction with policies "mitigating the long-term consequences of child poverty," with emphasis on younger children (birth to age 5). Early childhood interventions yield high returns: randomized-controlled experiments show spending on birth-to-five quality education programs for disadvantaged children produce a "7–13% annual return on investment."
  3. Policies must be broad based: beyond increasing parental income, support is crucial for education, access to health, adequate housing, and nutrition. Public investment in children's education is especially important given the pandemic's impact on skills accumulation.
- Specific policy-relevant observations:
  - Single parenthood and household size are structural drivers of child poverty; "single parents have more difficulties joining the labor market on a stable and full-time basis" and face higher in-work risk of poverty. In 2021, "19 percent of working single parents were at risk of poverty"—"twice the rate for other households with children."
  - Business cycle effects matter: changes in unemployment are "strongly associated with child poverty." The COVID-19 specific variable that performed best in regressions was the "stringency of containment measures," moderately associated with child poverty and significant for younger children, suggesting childcare closures affected parents’ capacity to work.

### Econometric investigation: drivers of child poverty (overview)
- Regression results (Table 4; robust to different specifications) indicate child poverty is associated with:
  - Structural features: demographics (household size, prevalence of single parenthood) and income inequality.
  - Economic cycle: notably labor market conditions and unemployment.
  - Social protection spending.
- The dummy for the GFC is highly significant, capturing additional crisis transmission channels.
- Among COVID-19-specific variables tested (closure of schools and nurseries, containment measures, social distancing), "the stringency of containment measures performs best" and is significant for younger children.

*IMF WORKING PAPERS Rising Child Poverty in Europe – Mitigating the scarring from the COVID-19 pandemic*

### 12.5 percent of households with dependent children. This was a 1.1 percentage point increase in a decade. The shares ran

### wpiea2023134-print-pdf - 12.5 percent of households with dependent children. This was a 1.1 percentage point increase in a decade. The shares ran

### Findings
- 12.5 percent of households with dependent children.
- This was a 1.1 percentage point increase in a decade.
- The shares range from

### Key statistics
- 12.5 percent
- 1.1 percentage point
- a decade

*Source: wpiea2023134-print-pdf - 12.5 percent of households with dependent children. This was a 1.1 percentage point increase in a decade. The shares ran*

### 3.3 percent in Slovenia and Greece to 29.3 percent in Estonia.

### wpiea2023134-print-pdf - 3.3 percent in Slovenia and Greece to 29.3 percent in Estonia.

### Key empirical findings (Severe material deprivation and correlates)
- Severe material deprivation across EU27 ranges from "3.3 percent in Slovenia and Greece to 29.3 percent in Estonia."
- Table 4 — Panel regressions (EU27, 2009–20) — highlights (selected coefficients, robust standard errors in parentheses; significance indicators preserved):
  - Stringency of containment during the pandemic: 0.161**** (0.043) for All Children (Baseline (1)); 0.203**** (0.051) for Younger Children (Baseline (1)).
  - GFC dummy: 0.188**** (0.054) and 0.247**** (0.063) in alternate specifications.
  - Gini coefficient: 0.101* (0.055) and 0.126** (0.060) in baseline specifications.
  - Average household size: 0.354**** (0.058); 0.379**** (0.076); 0.356**** (0.059).
  - Share of children with single parent (parents of respective age group): coefficients include 0.203** (0.080); 0.220*** (0.072); 0.285*** (0.108).
  - Unemployment rate: 0.229**** (0.043); 0.237**** (0.046); 0.236**** (0.048).
  - Share of children in households with very low work intensity (parents of respective age group): 0.174*** (0.062); 0.194*** (0.066); 0.160**** (0.082); 0.178**** (0.050).
  - Temporary employment share: 0.173* (0.090); 0.211** (0.102); 0.268* (0.139).
  - Family and children spending (share of GDP): -0.309**** (0.061); -0.282**** (0.062); -0.311**** (0.082); -0.276**** (0.081).
  - Other working age spending: -0.147 (0.119); -0.216 (0.137) (not statistically significant).
  - Observations: 322 and 310 (reported for specifications). R-Squared: 0.623, 0.627, 0.533, 0.539. Adjusted R-Squared: 0.582, 0.579, 0.478, 0.481. F-Statistic: 60.15****, 45.96****, 41.03****, 32.08****.
- Single parenthood is more strongly and more significantly associated with younger child poverty than with poverty of all children.
- Results are robust to an alternative business cycle measure (the output gap).

### Labor market mechanisms and pandemic effects
- Labor market participation is associated with child poverty: parents’ unstable or non‑full‑time participation limits stable labor earnings and contributes to child poverty.
- During the GFC, increases in low work‑intensity households were central to the transmission of crisis impacts to households with children, disproportionately affecting groups with low education, single parents, and migrants.
- During the COVID‑19 pandemic:
  - Job retention schemes prevented a surge in unemployment and a sharp decline in employment; the strong link between unemployment rate and child poverty suggests mitigation of child poverty increases.
  - Nonstandard workers (temporary contracts) experienced a stronger decline in employment and were typically less eligible for job retention schemes; this likely contributed to increased child poverty because:
    - Single parents rely more on nonstandard work than other parents (Figure 4).
    - The share of children in households with very low work intensity increased during the pandemic (Figure 5).
- At time of writing, severe material deprivation data for 2021 were not available; the increase in children living in households with low work intensity in 2021 is one reason to expect severe material deprivation increased further that year, even if unemployment declined. The share of children living with adults with a temporary contract was broadly similar in 2021 and 2020 (9.8 percent versus 9.8 percent).

### Childcare, female employment, and incentive structures
- Increasing parental employment (and working hours) is a key channel to reduce child poverty and mitigate long‑term scarring; policies include increasing employment rate, increasing net wage and predictability, and improving skills.
- Over the 2010s, enrollment of younger children in formal childcare increased by about 6 percentage points in the EU and was accompanied by a similar increase in female employment rate.
- Main impediment to further childcare use in the EU is financial cost (Figure 8).
- Childcare cost dynamics (selected points from figures and text):
  - Net and gross costs for full‑time center‑based childcare declined in recent years via tax cuts and benefits (Figure 9).
  - Childcare remains relatively more expensive for parents earning minimum wage than for parents earning average wage, though the gap narrowed between 2015 and 2021.
  - OECD (2020b) shows the effective tax rate on entering full‑time employment increases when childcare costs are accounted for in almost all 22 EU countries with data (Figure 10).
- Marginal effective tax rates (Figure 6) show that in several countries the total household income does not increase (or declines) when a parent works longer (marginal effective tax rate 100 percent or higher). But some countries (example: Slovak or Irish low earner single parent with two children under scenarios in Figure 6) can have negative marginal effective tax rates when increasing working time from 1/3 to 2/3 of full time.
- Barriers to parents working more include lack of flexible hours, inadequate parental leave (including paternity leave), limited leaves for child sickness, difficulties returning after parental leave, gender biases in employment, and access/cost of childcare.
- Despite increased access to childcare promoting female labor market participation, the impact on child poverty is muted because the poorest are less likely to use formal childcare: children under three in childcare centers are overrepresented in higher income quintiles; example from France: 30 percent of children in the least well‑off third benefit from formal childcare versus nearly 60 percent for the population as a whole.

### Policies to increase net wages and job stability
- Options to boost take‑home wage for poor parents:
  - Increase the minimum wage (can reduce child poverty but may affect employment opportunities for low‑skilled parents).
  - Reduce payroll taxes at the low end of the pay scale:
    - Cutting employee contributions raises net wage without increasing employer labor cost.
    - Cutting employer contributions can create low‑skilled jobs without reducing minimum wage (evidence cited).
  - Consider tax credits similar to the U.S. Earned Income Tax Credit: evidence shows increases in labor supply (notably single parents), reductions in poverty, and improvements in children’s health and human capital.
- Stable and predictable income reduces child poverty levels and mitigates long‑term impacts by reducing mental stress, improving parenting, improving infant brain activity (randomized evidence), and increasing school attendance over childhood and adolescence.
- Active labor market policies, structural and legal reforms, affordable childcare with adequate opening hours, and skills improvements increase the chance of stable employment.

### Skills and long‑term intergenerational considerations
- Parental educational attainment is strongly linked to children’s severe material deprivation (Figure 12); evidence points to the need for adequate and subsidized training for parents with the lowest educational level.
- Long‑term educational policies to reduce dropout rates and provide adequate work skills are essential to reduce intergenerational transmission of poverty.

### Fiscal policies: rapid reduction of child poverty and long‑term mitigation
- Fiscal spending on family and children (social protection spending on family and children as a share of GDP) is strongly negatively associated with children’s severe material deprivation (Table 4).
- Social transfers significantly reduce monetary poverty; their impact on children is larger than for the rest of the population (Figure 13).
- Fiscal measures have rapid effects on child poverty and are therefore crucial for rapid post‑COVID‑19 reduction in child poverty.
- Post‑GFC structural changes:
  - Spending on family and children contributed more than other social protection areas to post‑GFC fiscal consolidation; family and children spending declined in real terms in most EU countries between peak during the GFC and the trough post‑GFC.
  - Other social protection areas increased or were cut less; spending on old age and survivors was protected from cuts.
  - The share of family and children in total social protection spending declined from 2009 to 2013, raising concerns about underinvestment in children and unequal intergenerational distribution of austerity.
- Fiscal policy should balance rapid child poverty reduction (through transfers and family‑targeted spending) with public investment in physical infrastructure and human capital to mitigate long‑term scarring from pandemic‑related increases in child poverty.

*IMF staff calculations; wpiea2023134-print-pdf*

### 2013. See Table 5 for precise dates.

### wpiea2023134-print-pdf - 2013. See Table 5 for precise dates.

### Spending patterns and pandemic impact
- During the pandemic, the social protection system shielded the elderly better than children; the share of elderly suffering from severe material deprivation decreased during the pandemic, while it increased for children and the working age population.
- For the EU27:
  - Social protection spending on family and children increased by 9.2 percent in 2020; total social protection spending increased by 8.8 percent in 2020; total general government spending increased by 9.2 percent in 2020.
  - In real terms, total social protection benefits for family and children per inhabitant increased by 1.8 percent a year on average between 2008 and 2019.
  - In-kind benefits grew by 3.5 percent per year (2008–2019), compared with 0.9 percent annual increase for in-cash benefits.
  - Means-tested benefits increased by 2.9 percent per year (2008–2019) while non-means-tested benefits increased by 1.5 percent per year (2008–2019).
- In 2020, in-cash family and children spending per inhabitant grew by 7.2 percent in real terms — the largest such increase since the GFC period; in-kind spending growth in 2020 was 2.9 percent (below the 2008–19 annual average of 3.8 percent).

### Delivery and composition of benefits
- The share of social protection spending excluding family and children that is means-tested was at 9 percent in both 2008 and 2019; the share provided in kind stood, in both years, at 34 percent.
- The increase in family and children expenditure by 0.2 percent of GDP between 2008 and 2019 in the EU was entirely due to in-kind spending, driven by the increase of spending on childcare (Early Childhood Education and Care—ECEC).
- Post-GFC cuts in family and children spending were concentrated in in-cash spending; ECEC/in-kind spending was relatively protected.

### Regression evidence on effectiveness (key coefficients and findings)
- Breakdown results (summary findings):
  - Relying more on in-kind spending increases efficiency of family and children spending on poverty of younger children but reduces its impact on poverty of older children.
  - Combined impact on younger children’s poverty: in-kind spending -0.49 versus in-cash spending -0.34.
  - Combined impact for all children: in-cash spending -0.38 versus in-kind spending -0.24.
  - Non-means-tested spending has a larger impact on child poverty than means-tested spending (true for both in-cash and in-kind).
  - Means-tested in-kind spending has no statistically significant association with reduction in child poverty (not associated with a reduction).
- Selected regression coefficients from Table 6 (EU27, 2009–20; robust standard errors in parentheses):
  - Family and children spending: Total: -0.309**** (0.061)
  - Family and children spending: In cash (Means tested): -0.124*** (0.041)
  - Family and children spending: In cash (Not means tested): -0.258**** (0.043)
  - Family and children spending: In kind (Means tested): 0.011 (0.102) [not significant]
  - Family and children spending: In kind (Not means tested): -0.255** (0.099)
  - Younger children breakdown: Total: -0.311**** (0.082); In cash (Means tested): -0.099** (0.049); In cash (Not means tested): -0.239**** (0.048); In kind (Means tested): 0.035 (0.124); In kind (Not means tested): -0.526*** (0.174)
- Regression model fit and sample:
  - Observations: 322 (baseline), 310 (breakdown)
  - R-Squared: 0.623 (baseline), 0.657 (breakdown)
  - Adjusted R-Squared: 0.582 (baseline), 0.612 (breakdown)

### Policy recommendations and design implications
- Short-term priority: preserve current level of social protection spending on family and children; if fiscal space allows, prioritize increases targeted to reduce child poverty rapidly.
- Efficiency-enhancing reforms (options discussed):
  - Introduce a universal child benefit taxed at a sufficiently progressive and broad-based personal income tax so the benefit is larger for low-income parents; potential advantages: fiscal neutrality candidate, simpler administration, reduced stigma, minimized work disincentives.
  - If universal benefit is not adopted, consider:
    - Making part of in-cash transfers conditional on performance criteria (e.g., school attendance, health checks) to potentially alter parental behavior and increase investment in child development — but randomized trials show only marginal improvements in child health and education outcomes and no parental earnings effects.
    - Increasing transfers to specific high-need groups (parents of younger children, single parents). Example: Sweden’s experience suggests stable minimum income to single parents via minimum child support benefit can have positive impact; may need stronger enforcement of child alimony.
- Design cautions for means-testing and conditionality:
  - Pitfalls of means-testing: eligibility criteria may be too tight or too generous; some recipients neither poor nor entitled; low take-up due to stigma, complex/invasive procedures, or unawareness.
  - Means-testing needs careful design so withdrawal of benefits does not create work disincentives or undermine labor-market policies.

### Role of public investment to limit scarring from child poverty
- Complementarity: cash transfers and in-kind benefits are complementary — different channels and mechanisms; sustained, long-lasting transfers have larger long-term benefits.
- Recommended public investments:
  - Childcare infrastructure: accompany EU policies that reduce cost to parents with investments to increase availability of formal ECEC, notably in poorest areas, and to ensure quality of services (quality matters; low-quality childcare can be harmful).
  - Education: invest to improve inclusion of poor children, reduce isolation of disadvantaged students, recoup pandemic educational losses, and improve distribution of education financing; note that education spending at EU level declined by 0.4 percentage points of GDP to 4.7 percent of GDP between 2010 and 2019; capital spending on education averaged 0.4 percent of GDP during the period.
  - Health and nutrition: increase access to healthcare and medicine for poor children and poor pregnant women; provide adequate nutrition to poor children and pregnant women (FEAD is primarily a food aid scheme with 80 percent of its budget devoted to food).
  - Housing: ensure affordable adequate housing via social housing or financial support while avoiding creation of isolated poor neighborhoods that negatively affect child development.
- Evidence on long-term benefits of sustained transfers and investments:
  - Unconditional and predictable cash transfers reduce parental stress and improve parental involvement, leading to benefits into adulthood (e.g., better health, reduced risky behavior, improved educational and financial outcomes).
  - Literature documents positive long-term effects of early-life health and education investments.

### Costs, country examples, and headline statistics
- Germany: proposed “Basic Child Security” plan estimated cost about EUR 12 billion; met with resistance in part due to cost.
- Child poverty trends and impacts:
  - In 2020, the number of children suffering from severe material deprivation increased by 19 percent (0.9 million) in the EU.
  - Indirect evidence suggests further increases in 2021; the sharp increase in inflation in 2022 likely worsened the situation.
- FEAD (Fund for European Aid to the Most Deprived) characteristics:
  - FEAD primarily a food aid scheme: 80 percent of its budget devoted to food; children are important beneficiaries.
- Education spending: Between 2010 and 2019 education spending at the EU level declined by 0.4 percentage points of GDP to 4.7 percent of GDP; capital spending on education accounted for 0.4 percent on average during the period.

### Conclusion — policy priorities to limit scarring
- Two complementary short-term levers to reduce child poverty rapidly:
  - Deploy labor policies to increase parental employment, working hours, net wages, and labor income stability (noting some measures may take time to materialize).
  - Use social protection spending on family and children to achieve immediate impact; avoid repeating post-GFC cuts to these spending areas that would delay reduction in child poverty and increase scarring.
- Longer-term priorities:
  - Sustain social protection support and reform delivery to increase impact (consider universal child benefit or better-designed targeting/conditionality).
  - Scale public investment in childcare availability/quality, education inclusion and recovery, health and nutrition access for children and pregnant women, and affordable housing to reduce persistent intergenerational transmission of poverty.

*Source: IMF Working Paper — Rising Child Poverty in Europe – Mitigating the scarring from the COVID-19 pandemic (excerpts from the supplied content unit).*

### Annex I. Definition of Concepts

### Annex I. Definition of Concepts

### A. Survey Data

- European Union Statistics on Income and Living Conditions (EU-SILC)
  - The EU-SILC is a panel survey conducted in EU and other European countries whose micro-household data underpin both the Eurostat Income and Living Conditions (ILC) database as well as the OECD Income Distribution Database.
  - The survey provides both cross-sectional data pertaining to a given time or time-period as well as longitudinal data pertaining to individual-level changes over time, observed periodically or over a four-year period.
  - Comparability issues posed by differences in data collection across countries (for example, including reliance on household surveys as opposed to ‘register’ data) as well as the allowance for different concepts of self-employment income are addressed in the survey by conceptual harmonization of target variables and the so called “ex ante output harmonization model” employed by Eurostat (Eurostat 2007).
  - Limitations include:
    - exclusion of social transfers in kind from disposable income;
    - exclusion of capital gains;
    - restriction of the data to the population living in private households.

- Labor Force Survey (LFS)
  - LFS data in the time period relevant to this working paper are considered highly comparable across countries and time given the EU-LFS use of output harmonization.
  - Increased content stability and frequency of surveys have also been implemented from 1998 to counterbalance allowances for country-specific surveys.

- European Social Survey (ESS)
  - The ESS employs stringent sampling and collection design, data processing, and quality assessment checks to address measurement challenges for concepts such as preference or attitude.
  - Quality procedures include periodic reports on measurement quality and equivalence of survey responses, frequent nonresponse bias analyses, response rate floors, and monitoring and recording of contextual data taken from national media.

### B. Measuring Inequality

- Equivalized Disposable Income
  - Definition: total income of a household available for spending or saving, divided by the number of household members converted into equivalized adults by Eurostat.
  - Exclusions from calculations: people with missing values for equivalized disposable income and those living in collective households and in institutions.
  - Equivalence scale (Eurostat 2014):
    - the first household member aged 14 years or older as 1 person
    - each other household member aged 14 years or older as 0.5 person
    - each household member aged 13 years or younger as 0.3 person
  - For regional time-series analysis, data are converted into 2013 prices by deflating by national consumer price indices, indexed to 100, then weighted by PPP before aggregation into their respective regions.

- Gini Coefficient
  - The Gini coefficient used refers to the Gini of equivalized disposable income.
  - Advantages: independent of the sample mean and population size, symmetrical, and sensitive to transfers of income from the top to the bottom of a distribution.
  - Limitation: not decomposable or additive across subgroups like age, and thus insufficient for more granular subgroup analysis.

### C. Measuring Poverty

- Material Deprivation
  - Material Deprivation aggregates various dimensions of non-monetary well-being into a single measure to capture an individual’s “command over resources”.
  - Severe material deprivation rate definition: the enforced inability (rather than the choice not to do so) to pay for at least four of the following nine items considered by most people to be desirable or even necessary to lead an adequate life:
    - to pay their rent, mortgage or utility bills
    - to keep their home adequately warm
    - to face unexpected expenses
    - to eat meat or proteins regularly
    - to go on holiday (a week’s annual holiday away from home)
    - a television set
    - a washing machine
    - a car
    - a telephone

- At-Risk-of-Poverty Thresholds and Rates
  - Relative at-risk-of-poverty threshold: 60 percent of the national median equivalized disposable income after social transfers (unless otherwise indicated).
  - At-risk-of-poverty rate: proportion of persons with an equivalized disposable income below that threshold.
  - Subgroup calculations: where figures for subgroups exist, they are calculated based on the poverty threshold for the entire population.
  - Persistent at-risk-of-poverty rate: percentage of the population living in households where the equivalized disposable income was below the at-risk-of-poverty threshold for the current year and at least two out of three of the preceding years.
  - Notes on relative measures:
    - These are measures of relative poverty and can obscure effects of changes to the poverty threshold when the living standards of the entire population change.
    - Example: after the financial crisis, an overall decline in national income led to a decrease in the median income, allowing relative poverty to remain somewhat stable and masking the absolute deterioration of living conditions.
  - Conceptual concern: sole use of income as a proxy for living standards omits other assets and wealth, presenting an incomplete picture of economic well-being.
  - Absolute poverty approach: anchoring the poverty threshold (anchored at-risk-of-poverty rate) fixes the cutoff level over time (except for inflation adjustments) to isolate development of poverty from shifting poverty lines.
  - European-level indicator: “at risk of poverty and social exclusion” sums persons who are either at risk of poverty, severely materially and socially deprived, or living in a household with a very low work intensity; people are included only once even if in more than one of these situations.

*Annex I. Definition of Concepts — IMF Working Paper*

### 2020. Both pooling, fixed effects, and random effects specifications have been tested. The pooling and Hausman

### Content unit: wpiea2023134-print-pdf - 2020. Both pooling, fixed effects, and random effects specifications have been tested. The pooling and Hausman

### Estimation approach and diagnostics
- Pooling, fixed effects, and random effects specifications were tested; pooling and Hausman tests led to selection of the fixed effects model.
- Heteroscedasticity was established; a heteroscedasticity-consistent estimation of the covariance matrix was performed.
- Checks performed: serial correlation, unit roots, and multicollinearity.

### Data definitions, coverage, and sources
- Period covered: 2009–2020.
- Cutoff date for the data: November 22, 2022.
- Main data source: Eurostat; COVID-19 impact variables taken from IMF reports or other publicly available sources.
- Dependent variables (examples):
  - sevmatdepr18 — ilc_mddd11: Severe material deprivation rate of population below 18 measured as proportion of the children below 18 years of age.
  - sevmatdepr6 — ilc_mddd11: Severe material deprivation rate of children below 6 measured as proportion of the children below 6 years of age.
  - rop18pc — ilc_li02: Proportion of children below 18 years of age at risk of poverty.
  - anrop18pc — ilc_li22: At-risk-of-poverty rate anchored at a fixed moment in time (2005) for children below 18 years.
- Explanatory variables (selected):
  - gini — ilc_di12: Gini coefficient of equivalized disposable income - EU-SILC survey.
  - hhdsize — ilc_lvph01: Average household size - EU-SILC survey.
  - singlech2 — ilc_lvps20: Share of children (aged less than 18) living with a single parent - EU-SILC survey.
  - childcare3 — ilc_caindformal: Share of children below 3 years of age not in any form of formal childcare or education - EU-SILC survey.
  - eduatt02parents18 — lfst_hhacednc: Share of children below 18 years of age, whose parents are with less than primary, primary and lower secondary education (ISCED11 levels 0-2).
  - gdpppspc — nama_10_pc: GDP in current prices, purchasing power standard (PPS, EU27 from 2020) per capita.
  - unempl — une_rt_a: Unemployment rate.
  - lowworkint18 — ilc_lvhl11: Children below 18 years of age living in households with very low work intensity (percent of total population aged less than 60).
  - familyben1 — gov_10a_exp: Social protection expenditures of the general government for “family and children” in percent of GDP.
- Crisis and COVID indicators (selected):
  - gfc: Dummy for Global financial crisis and sovereign debt crisis (2009-13).
  - stringency: Stringency of containment during the pandemic – Oxford COVID-19 Government Response Tracker.
  - covid_cases_pa_pmln: Number of COVID-19 cases per annum per million of population.
  - covid: Number of days either in partial opening or full closure of schools due to COVID-19.
  - fiscalsupportgdp: Fiscal support during COVID-19 (percent of GDP) – IMF October 2020 Fiscal Monitor Report database.
  - onbudgetgdp: On-budget fiscal support during COVID-19 (percent of GDP) – IMF October 2020 Fiscal Monitor Report database.
  - outputgap: Output gap (WEO database – Extracted on September 28, 2022).

### Data transformations
- All indicators (except dummy variables) are taken in relative terms (percent of GDP, percent of total population, etc.).
- All data (except dummy variables) standardized by subtracting the mean and dividing by the standard deviation.
- Logistic transformation applied to dependent variables: ln[y/(1−y)], since child poverty indicators are proportions ∈[0,1].

### Robustness checks and alternative specifications
- Results are robust to alternative specifications, methods (machine learning), and to using at-risk-of-poverty rate instead of severe material deprivation.
- Alternative cyclical indicator: output gap used instead of unemployment rate; cyclical conditions remain significant. Over the period, unemployment rate and output gap correlation: -0.68.
- Educational level of parents is significantly associated with child poverty in some tests but becomes insignificant when added to baseline regressions (impact likely captured by other variables); machine learning confirms importance of parental education.
- Coverage robustness: baseline regressions robust to expansion to 38 European countries (EU27 plus 11 others), but missing data prevented full presentation.

### Key panel regression results (selected tables and coefficients)
- Table II.2 — Streamlined Specification — Severe Material Deprivation
  - All Children (Less than 18):
    - Output gap: -0.085**** (0.043)
    - Share of children with single parent: 0.377*** (0.130)
    - Family and children spending: -0.410**** (0.095)
    - Observations: 320
    - R-Squared: 0.552
    - Adjusted R-Squared: 0.502
    - F-Statistic: 58.90****
  - Younger Children (Less than 6):
    - Output gap: -0.112**** (0.028)
    - Share of children with single parent: 0.413**** (0.076)
    - Share of children living in households with very low work Intensity: 0.270*** (0.049)
    - Family and children spending: -0.425**** (0.117)
    - Observations: 320
    - R-Squared: 0.439
    - Adjusted R-Squared: 0.376
    - F-Statistic: 37.40****

- Table II.3 — At Risk of Poverty — Panel Regressions (selected columns)
  - Dependent variable: Anchored at-risk-of-poverty rate (Less than 18)
    - GFC dummy: 0.114 (0.087)
    - Gini coefficient: 0.275** (0.134)
    - Average household size: 0.366** (0.153)
    - Share of children living in households with very low work intensity: 0.383**** (0.072)
    - Temporary employment share: 0.389** (0.163)
    - Family and children spending: -0.308*** (0.117)
    - Observations: 275
    - R2: 0.666
    - Adjusted R2: 0.625
    - F Statistic: 60.77****
  - Dependent variable: At-risk-of-poverty rate (Less than 18)
    - Average household size: 0.081 (0.116)
    - Share of children with single parent: 0.135** (0.064)
    - Unemployment rate: 0.061 (0.060)
    - Share of children living in households with very low work intensity: 0.310*** (0.095)
    - Temporary employment share: 0.340** (0.133)
    - Family and children spending: -0.256*** (0.094)
    - Observations: 323
    - R2: 0.459
    - Adjusted R2: 0.398
    - F Statistic: 35.09****
  - Dependent variable: At-risk-of-poverty rate (Less than 6)
    - Family and children spending: -0.207 (0.165)
    - Share of children living in households with very low work intensity: 0.504**** (0.099)
    - Temporary employment share: 0.269** (0.166)
    - Observations: 323
    - R2: 0.354
    - Adjusted R2: 0.281
    - F Statistic: 22.65****

(Note: Robust standard errors in parentheses, **** p<0.001, *** p<0.01, ** p<0.05, * p<0.10. 1/ For parents of the respective children age group.)

### Machine learning (ML) alternative estimation — models, performance, and findings
- Six ML models used to identify drivers of severe material deprivation for children below 18 (SMD18) and below 6 (SMD6).
- Table II.4 — Comparison of Machine Learning Models’ Performance
  - All children — Mean Squared Error / R2:
    - Linear regression: 0.068 / 0.919
    - Elastic net: 0.063 / 0.926
    - K-nearest neighbors: 0.047 / 0.945
    - Support vector regression: 0.051 / 0.939
    - Random forest: 0.054 / 0.937
    - Extreme Gradient Boosting: 0.056 / 0.934
  - Small children — Mean Squared Error / R2:
    - Linear regression: 0.115 / 0.863
    - Elastic net: 0.102 / 0.879
    - K-nearest neighbors: 0.084 / 0.900
    - Support vector regression: 0.095 / 0.887
    - Random forest: 0.082 / 0.902
    - Extreme Gradient Boosting: 0.095 / 0.887
- ML main findings:
  - Models show good generalization and prediction on test data with R2 around 0.9 and higher.
  - Highest-impact determinants (averaged across models): household characteristics and income, labor market participation, and family and children benefits.
  - Additional important determinants highlighted by ML: low childcare enrollment and low parental education.
  - Model-specific highlights:
    - Elastic net and k-nearest neighbors: average household size has the highest contribution.
    - Support vector machine: GDP per capita PPS identified as main determinant.
    - Random forest: share of children below 3 years not in formal childcare or education identified as most important factor.
- Interpretations from figures and plots:
  - Income level (GDP) and social benefits negatively associated with child poverty.
  - Household size, unemployment, income inequality, share of households with very low work intensity, and lower childcare enrollment positively associated with child poverty.
  - Random forest indicates threshold effects:
    - Child poverty declines until GDP per capita above 28,000 PPS; beyond that point, increases in income have little effect.
    - Significant reduction of child poverty when family and children benefits reach around 3% of GDP.

### ML technical notes and implementation details
- Same data standardization and logistic transformation applied as in panel regressions.
- Dataset split: 75 percent train, 25 percent test.
- Hyperparameters tuned via cross-validated grid-search over predefined parameter grids.
- Cross validation: dataset split into ‘k’ subsamples; repeated procedure and averaged outcomes reported.

### Summary of substantive empirical conclusions
- Family and children spending (benefits) are consistently negatively associated with child poverty across specifications and measures (e.g., coefficients: -0.410**** (0.095) in Table II.2; -0.308*** (0.117) in Table II.3 anchored rate).
- Labor market conditions are important drivers:
  - Unemployment/output gap matter for child poverty (output gap coefficients: -0.085**** (0.043) and -0.112**** (0.028) in streamlined specification).
  - Low work intensity and temporary employment share are positively associated with higher child poverty (e.g., low work intensity 0.383**** (0.072) in anchored at-risk-of-poverty).
- Household structure and demographics:
  - Larger average household size associated with higher child poverty (e.g., 0.366** (0.153) in anchored at-risk-of-poverty).
  - Share of children living with single parent associated with higher child poverty in several specifications (e.g., 0.377*** (0.130) in Table II.2).
- Income inequality (Gini) positively associated with at-risk-of-poverty (e.g., 0.275** (0.134) in anchored at-risk-of-poverty).
- Robustness: results hold across alternative definitions of poverty, alternative cyclical measures, alternative country coverage, and machine learning approaches.

*Source: IMF staff calculations, from the IMF Working Paper unit described in the supplied content.*

### 2. Elastic Net

### 2. Elastic Net

### Elastic Net: definition and loss function
- The elastic net is an extension of the linear regression.
- Ordinary least squares estimation minimizes the sum of the squares of the errors: ∑(y_i − ŷ_i)^2_{n i=1}.
- The elastic net imposes two penalties on this model, each weighted by the hyperparameter α ∈ [0,1]:
  - L1 penalty: ∑|β_j|_{p j=1}
  - L2 penalty: ∑β_j^2_{p j=1}
- Elastic net penalty (as presented):
  - Elastic net penalty = α * L1 + (1 − α) * L2 = α ∑β_j^2_{p j=1} + (1 − α) ∑|β_j|_{p j=1}
- Loss function of the elastic net model (as presented):
  - Loss function = ∑(y_i − ŷ_i)^2_{n i=1} + λ [ α ∑β_j^2_{p j=1} + (1 − α) ∑|β_j|_{p j=1} ]

### Notes on penalties
- L1 penalty: ∑|β_j|_{p j=1}
- L2 penalty: ∑β_j^2_{p j=1}

### Random Forest Regression
- Random forest is an ensemble method aggregating multiple decision trees to improve joint performance.
- Individual decision trees are classifiers organized with root, branches, internal and leaf nodes.
- Single decision trees are prone to bias and overfitting; a random forest constructed from unrelated decision trees can address these issues.
- Random Forest models can capture complex models and, due to their tree structure, allow for straightforward interpretation.

### Extreme Gradient Boosting (XGBoost)
- XGBoost (eXtreme Gradient, Chen and Guestrin 2016) builds on random forest by adding new trees one by one to correct prediction errors made by existing ones.
- XGBoost includes built-in algorithms allowing faster execution via parallelization and other techniques.

### Support Vector Regression
- Support vector machine maps observations through a kernel transformation (usually to higher dimensions) to separate observations with a maximal gap.
- For regression, the SVM fits the best line within a threshold value (the distance between the hyperplane and the boundary line).
- Support vector regression characteristics:
  - Good generalization capacity and high prediction accuracy.
  - Robust to outliers.
  - Reported to be less effective for large datasets and performs poorly when the dataset has more noise.

### K-Nearest Neighbors (KNN)
- K-Nearest Neighbors is used for classification and regression; it uses proximity for prediction.
- Training uses the entire dataset; predictions use the mean or median of the k-most similar observations.
- Distance measures for continuous variables: Euclidean, Manhattan, or Minkowski distance.
- Choice of k is important; typically, the larger the k, the more accurate the model.
- Method is prone to overfitting and scaling-related issues.

### Interpretation of Results: overview
- Few machine learning methods are straightforward to interpret; historically treated as "black boxes."
- Interpretation techniques fall into two categories:
  - Summary-based: insights about average contribution of features to the outcome.
  - Instance-based: breakdown for a specific observation.
- The main explanatory methods summarized: Permutation Feature Importance, Partial Dependence Plots (PDP), LIME, SHAP.

### 1. Permutation Feature Importance
- Purpose: estimate impact of random permutation of a single variable's values on the model score, independent of model type.
- Algorithmic steps:
  (i) Calculation of the original score of the model (i.e., R^2 or mean squared error).
  (ii) Permutation of the values of each feature and calculation of the model score based on permuted values. This step might be repeated multiple times and an average score might be calculated.
  (iii) Comparison of model scores based on original and permuted datasets. The worse the model score after permutation, the more important the respective feature.
- Caveat: with high correlation between features, importance of each correlated feature will appear lower because other features can still explain the outcome.
- Permutation importance can be calculated on training or test subsamples; the authors used the test dataset to assess contribution to generalization power and check for overfitting.

### 2. Partial Dependence Plots (PDP)
- PDPs estimate the average prediction of the outcome based on different values of one or more selected explanatory variables, conditional on the rest of the features.
- PDPs depict only average marginal effects and may hide heterogeneous relationships due to interactions.
- PDPs take the average of other variables and explicitly assume no correlation between explanatory variables.
- Value added: visualization of relationship type between outcome and selected feature (linear, monotone, discrete jumps, etc.).

### 3. Local Interpretable Model-agnostic Explanations (LIME)
- LIME approximates a complex (black box) model locally with an interpretable model (linear regression or decision tree).
- Algorithmic steps:
  (i) Selection of a specific observation.
  (ii) Perturbation of the dataset and calculation of predictions for new data points.
  (iii) Assignment of weights according to proximity of the data points to the selected observation.
  (iv) Training of a weighted interpretable model on the new dataset.
  (v) Interpretation of the locally approximating model.
- LIME provides only local explanations and depends strongly on the adopted definition of a for the instance and other parameter values.

### 4. SHapley Additive exPlanations (SHAP)
- SHAP derives from Shapley values in coalition game theory and calculates contribution of each feature value to prediction minus the average prediction for all instances.
- Shapley values are the average marginal contribution of a feature value taken across all possible coalitions (all possible values of other features).
- Theoretical properties satisfied:
  (i) Efficiency—the sum of the feature contributions adds up to the difference of the prediction for the feature value at this instance and the average.
  (ii) Symmetry—if two features contribute equally to all possible coalitions, their Shapley values are the same.
  (iii) Dummy—if a feature does not change the predicted value in all possible coalitions, it has a Shapley value of 0.
  (iv) Additivity—for the Shapley value for an aggregated object is the sum of the Shapley values of its components.
- Disadvantages: computational intensity and need to have access to all features and all data even after model estimation.
- SHAP (Lundberg and Lee 2017) provides both local and global breakdowns by feature contributions.
- The authors display a summary plot where all features and observations are shown and (in python) features are ordered according to their importance.

*IMF Working Papers — chapter excerpt: 2. Elastic Net (and subsequent ML methods and interpretability techniques) from the provided PDF content.*

### References

### References

### Major thematic areas covered by the references
- Multidimensional poverty and inequality measurement and theory (e.g., Aaberge and Bandolini (2015); Sen (1992), (1999)).
- Early childhood development, human capital formation, and lifelong impacts of childhood circumstances (e.g., Almond and Currie (2011); Currie (2001), (2009); Heckman (2008)).
- Intergenerational transmission of inequality and mobility (e.g., Aizer and Currie (2014); Corak (2013); Bolt et al. (2021)).
- Child poverty measurement, material deprivation, and distributive implications of family policies (e.g., Boarini and Mira d'Ercole (2006); Förster et al. (2013); Thévenon et al. (2018)).
- Impacts of COVID-19 on education, learning losses, and labor markets (e.g., Agostinelli et al. (2022); Engzell, Frey, and Verhagen (2021); Werner and Woessmann (2021); Fuchs‑Schündeln et al. (2020)).
- Cash transfers, child tax credits, and income-support policies for children and families (e.g., Barr, Eggleston, and Smith (2022); Berkowitz et al. (2022); Dreyer (2022); Pulliam and Reeves (2021)).
- Childcare, parental labor force participation, and policies to reduce child poverty (e.g., Hufkens et al. (2019); Morrissey (2017); Borowsky et al. (2022); Hermes et al. (2023)).
- Education policy, early childhood programs, and economic returns to preschool interventions (e.g., Conti, Heckman, and Pinto (2016); Garcia et al. (2016); Duncan and Magnuson (2013)).
- Health, stress, and child development (e.g., Blair (2010); Shonkoff et al. (2012); Hair et al. (2015)).
- Comparative and international policy frameworks (European Commission publications; OECD reports; UNICEF reports; World Bank (2015)).

### Representative influential works and policy-focused publications (selected citations with exact bibliographic entries as given)
- Aaberge, Rolf and Andrea Bandolini (2015). “Multidimentional Poverty and Inequality,” in A B. Atkinson and François Bourguignon, Handbook of Income Distribution, vol. 2, Amsterdam: Elsevier, pp. 141–216.
- Agostinelli, Francesco, Matthias Doepke, Giuseppe Sorrenti, and Fabrizzio Zilibotti (2022). “When the Great Equalizer Shuts Down: Schools, Peers, and Parents in Pandemic Times,” Journal of Public Economics, vol. 206, 104574.
- Almond, Douglas and Janet Currie (2011). “Human Capital Development Before Age Five,” in David Card and Orley Ashenfelter (eds.) Handbook of Labor Economics, vol. 4 Part B, Amsterdam: Elsevier, pp. 1315–1486.
- Barr, Andrew, Jonathan Eggleston, and Alexander A. Smith (2022). “Investing in Infants: The Lasting Effects of Cash Transfers to New Families,” The Quarterly Journal of Economics, 45 p.
- Engzell, Per, Arun Frey, and Mark D. Verhagen (2021). Learning Loss due to school closure during the COVID-19 Pandemic, Proceedings of the National Academy of Sciences of the United States of America, Vol. 118(17), 7 p.
- Heckman, James J. (2008). “School, Skills, and Synapses,” Economic Inquiry, Vol. 46(3) July, pp. 289–324.
- Hufkens, Tine, Francesco Figari, Dieter Vandelannoote, and Gerlinde Verbist (2019). Investing in Subsidized Childcare to Reduce Child Poverty: An Adequate Strategy. Seville: European Commission, Joint Research Center, JRC Working Papers on Taxation and Structural Reforms Mo. 06/2019, 20 p.
- IMF (2021). A Fair shot, Washington, DC: IMF: Fiscal Monitor April 2021, 105 p.
- OECD (2018). Poor children in rich countries: why we need policy action, Paris: OECD, Policy Brief on Child Well-Being, 8 p.
- Shonkoff, Jack P., Andrew S. Garner, Benjamin S. Siegel, Mary I. Dobbins, Marian F. Earls, Laura McGuinn, John Pascoe, and David L. Wood (2012). “The Lifelong Effects of Early Childhood Adversity and Toxic Stress,” Pediatrics, vol. 129(1), e232–e246.
- Troller-Renfree, Sonya V., Molly A. Costanzo, Greg J. Duncan, Katherine Magnuson, Lisa A. Gennetian, Hirokazu Yoshikawa, Sarah Halpern-Meekin, Nathan A. Fox, and Kimberly G. Noble (2022). “The impact of a poverty reduction intervention on infant brain activity,” Proceedings of the National Academy of Sciences of the United States of America, vol. 119(5).
- World Bank (2015). “Early childhood development,” World Development Report, pp. 98–110.

### Methodology, data, and computational tools cited
- Machine learning and interpretable models referenced: Chen and Guestrin (2016) “Xgboost: A scalable tree boosting system,” Lundberg and Lee (2017) “A unified approach to interpreting model predictions,” Molnar (2018) InterpreTable Machine Learning, and Müller and Guido (2016) Introduction to Machine Learning with Python.
- Empirical and cross-country measurement approaches: Eurostat (2007) Comparative EU statistics on Income and Living Conditions; Förster et al. (2013) The OECD Approach to Measure and Monitor Income Poverty across Countries; Chen et al. (2018) Inequality and Poverty Across Generations in the European Union (IMF Staff Discussion Note 18/01, 51 p).

### Policy-relevant emphases reflected across the references
- Targeting versus universal approaches to child and family support (e.g., Gugushvili and Hirsch (2014); Brady and Burroway (2012); OECD and European Commission policy documents).
- The role of cash transfers, child tax credits, and in-kind services (childcare, early childhood education) in reducing child poverty and improving long-term outcomes (multiple empirical and policy analyses cited).
- Mitigating learning losses and scarring from the COVID-19 pandemic through focused education and social policy interventions (numerous COVID-era studies and policy notes cited).
- Gender and single-parent household dynamics in poverty risk and policy design (e.g., Nieuwenhuis (2017, 2020, 2021); Heine (2016); Le Pape et al. (2015)).

*Rising Child Poverty in Europe - Mitigating the scarring from the pandemic, Working Paper No. WP/2023/134 — References*

---


_Source: https://www.imf.org/-/media/files/publications/wp/2023/english/wpiea2023134-print-pdf.pdf_