Measuring human capital using global learning data

The study addresses how best to measure human capital—traditionally proxied by years of schooling—to understand its role in economic development. Evidence shows schooling does not always translate into learning, especially in low- and middle-income countries, creating a gap between enrolment/attainment and actual cognitive skills. International assessments (e.g., PISA, TIMSS) have expanded but underrepresent developing countries, limiting global comparability. The authors aim to bridge this gap by constructing a Harmonized Learning Outcomes (HLO) database covering 164 countries (98% of the global population) from 2000–2017, enabling cross-country and over-time comparisons of learning. They then use these measures to reassess the contribution of human capital to income differences and growth, hypothesizing that learning-based measures will better explain economic outcomes than schooling-based proxies.

Earlier research often measured human capital via schooling (years of education, enrolment rates), including in influential indices like the UN’s Human Development Index. However, schooling frequently fails to predict learning, especially in developing contexts, and aggregate returns to schooling alone can be weak. Studies have suggested that when human capital is measured by learning (cognitive skills), it is more strongly associated with growth. Development accounting literature has produced wide-ranging estimates for the contribution of human capital to income differences—ranging from nearly all to potentially none—owing to challenges in measuring the quality dimension (e.g., reliance on Mincerian returns, immigrant returns, and skill premia, each requiring strong assumptions). Recent efforts have improved cross-country learning data, but with limited coverage outside high-income countries. The HLO database builds on this literature by providing broader coverage and a direct measure of learning quality suitable for development accounting and growth analyses.

Data construction (HLO database): The World Bank identified and collated student assessment data worldwide, incorporating seven regimes: three international assessments, three regional standardized assessments (covering much of sub-Saharan Africa and Latin America), and the Early Grade Reading Assessment (EGRA). Coverage spans 164 countries from 2000–2017 (2,023 country-year observations), disaggregated by schooling level (primary, secondary), subject (reading, mathematics, science), and gender, and includes mean scores and standard errors. Disaggregation by gender is available for 98.5% of observations. The database excludes pre-2000 data due to quality concerns.

Harmonization/linking: The authors link regional assessments to international assessments within the same subject, schooling level, and adjacent years to place all scores on a common scale (HLO). International assessments include PISA and TIMSS; regional assessments include LLECE and PASEC, among others. The linking procedure adjusts for systematic differences between tests to produce comparable HLO scores for countries participating only in regional assessments. A high-performance benchmark is set at TIMSS 625; a low-performance benchmark at 300 (aligned with minimum regional benchmarks). Psychometric principles of test equating/scaling/linking underpin the approach.

Descriptive and trend analyses: Average HLO scores are computed by country over 2000–2017 and summarized by region. To contrast quantity vs quality, expected years of schooling are compared with primary HLO scores for the most recent year available. Time trends in enrolment vs learning use adjusted primary enrolment ratios and primary HLO scores for 2000–2015, restricting to countries with at least two data points. Regional time trends are estimated with country-fixed-effects regressions: y = α_r + β_r t + δ_rc, where t is year and δ_rc are country dummies within region r; this mitigates composition bias over time. Additional analyses condition on enrolment and report coefficients/p-values.

Development accounting: The authors construct human capital stocks combining schooling and learning. They incorporate a rate-of-return parameter to the learning component (w) based on microeconomic evidence, conducting sensitivity analyses with w = 0.15, 0.20, 0.25; w = 0 implies schooling-only. They compute several standard decompositions used in the literature to quantify the contribution of human capital to cross-country income differences: h0/h1, (h0/h1)/(y0/y1), var(log h)/var(log y), and (ln h0 − ln h1)/(ln y0 − ln y1), where subscripts denote percentiles (e.g., 90th vs 10th). The accounting uses 131 countries; schooling data from prior literature, GDP per worker from Penn World Tables v9.0, and learning from HLO.

Growth regressions: To assess predictive power for growth, they regress the average annual growth rate of GDP per capita (2000–2010) on different human capital measures (HLO-based, Penn World Tables human capital index, schooling-only measures, HDI education index), controlling for log initial GDP per capita (year 2000). All independent variables are log-transformed for comparability. The sample includes 107 countries, excluding those in civil war, inflation crises, or with natural resource rents above 15%. They report coefficients, standard errors, p-values, R^2, and run specifications with alternative combinations of human capital measures and controls.

• Coverage and comparability: The HLO database spans 164 countries (98% of global population), including two-thirds developing economies, disaggregated by subject, schooling level, and gender, with standard errors for uncertainty.
• Global learning patterns: Average regional HLO scores (2000–2017) show substantial gaps: East Asia & Pacific 445; Europe & Central Asia 489; Latin America & Caribbean 402; Middle East & North Africa 399; North America 529; sub-Saharan Africa 342; South Asia 335. High-income countries outperform developing regions. Examples: Japan 553 vs United States 521; Kenya 444 and Tanzania 416 comparable to Latin America (e.g., Mexico 435); India 368 similar to Uganda 369; Chile 449 comparable to some European countries (e.g., Georgia 437).
• Schooling vs learning: Many countries achieve high expected years of schooling but low learning (e.g., Brazil 11.7 years with HLO 426; Ghana 11.6 years with HLO 229). Over 2000–2015, enrolment rose markedly (e.g., MENA: 95% to 99%), but learning stagnated (e.g., around 380). A panel regression of primary learning on primary enrolment with country fixed effects yields a non-significant association (coefficient −0.247; p = 0.673), indicating higher enrolment does not translate into higher measured learning on average.
• Development accounting: Incorporating learning alongside schooling implies human capital explains roughly one-fifth to about half of cross-country income differences, depending on the assumed rate of return to learning (w ∈ {0.15, 0.20, 0.25}). These estimates lie between extremes in prior literature (zero to nearly all). Substantial heterogeneity emerges: human capital explains a much larger share of income differences in advanced economies than in sub-Saharan Africa, and more in high-income than low-income country groups.
• Growth linkages (2000–2010): HLO-based human capital is more strongly and robustly associated with growth than alternatives. Example coefficients: ~0.072 (SE 0.018, p < 0.001) and ~0.059–0.069 across specifications with controls. By contrast, Penn World Tables human capital, schooling-only measures, and the HDI education index show weaker and often insignificant relationships once controls are included. Models control for initial income and show higher R^2 when including HLO measures.

The HLO database directly addresses the measurement challenge underlying human capital’s role in development by adding a quality (learning) dimension with near-global coverage. The descriptive analyses reveal a global learning crisis: rapid gains in schooling access have not been matched by gains in learning, particularly in developing regions. The development accounting results reconcile prior divergent estimates by showing that, when learning is incorporated, human capital explains a meaningful but not dominant share (about 20–50%) of income differences. This aligns with broader models where human capital interacts with other growth drivers. The growth regressions demonstrate that learning-based human capital has a stronger and more consistent association with subsequent growth than schooling-based proxies or composite indices, underscoring the importance of measuring skills rather than years alone. Heterogeneity across income groups and regions further shows that extrapolating from high-income samples can mislead; incorporating countries at different development stages is essential for accurate inference and policy targeting.

The paper introduces the Harmonized Learning Outcomes (HLO) database—one of the largest, most current global datasets on learning—enabling comparable measurement of human capital quality across 164 countries from 2000–2017. Using this database, the authors document limited global progress in learning despite substantial enrolment gains, quantify human capital’s contribution to income differences at about one-fifth to one-half when learning is included, and show that HLO-based human capital more strongly predicts economic growth than standard schooling-based measures (e.g., Penn World Tables human capital index, HDI education index). The work supports a shift in focus from schooling attainment to learning outcomes for policy and monitoring (e.g., SDGs, World Bank Human Capital Index). Future research should expand participation in assessments (particularly in low- and middle-income countries), improve linking and equating methods, refine human capital aggregation (including health and non-cognitive skills), and update the HLO database regularly to track progress and evaluate policy impacts.

• Data coverage and quality: Learning data before 2000 are limited/low quality; some countries have sparse time-series coverage. Although the HLO includes standard errors and fixed-effects analyses to mitigate biases, residual measurement error and uncertainty remain.
• Harmonization assumptions: Linking regional to international assessments relies on equating assumptions within subjects, levels, and adjacent years. Differences in test design, proficiency distributions, and participation may affect comparability despite adjustments.
• Selection effects: As enrolment expands, newly included lower-performing students may depress average scores (selection). While trends persist even where enrolment is stable and fixed effects are used, selection cannot be fully ruled out.
• Parameter sensitivity in accounting: Development accounting relies on assumed returns to learning (w) and on functional forms used in decompositions; results vary with these choices.
• Sample exclusions in growth regressions: Countries in conflict/crisis and resource-dependent economies are excluded, which may affect external validity. Growth models may also omit other determinants (institutions, technology) correlated with human capital.
• Scope of human capital: The HLO focuses on cognitive skills (reading, math, science). Other dimensions (health, socio-emotional skills) are not incorporated but may influence income and growth.