Better skills data for smarter financing of education and training

Annex A. Types of data sources for decision making

Population and household surveys

Cross-sectional household and individual surveys

Cross-sectional household and individual surveys are widely used instruments for understanding the educational, social and economic characteristics of a population. These surveys collect information at a single point in time from nationally or regionally representative samples, providing a detailed snapshot of key elements such as access to education, levels of educational attainment, patterns of school attendance, reasons for dropout, household socio‑economic characteristics, living conditions and demographic profiles. They constitute a foundational source for describing system performance and identifying disparities across groups and territories.

Examples of well-established cross-sectional household surveys include the Living Standards Measurement Study (LSMS) series of the World Bank, Demographic and Health Surveys (DHS) and UNICEF’s Multiple Indicator Cluster Surveys (MICS). Many countries also operate national cross-sectional household surveys that integrate education modules, such as CASEN in Chile, ENIGH in Mexico, GEIH (household module) in Colombia, and Enquête Revenus Fiscaux et Sociaux in France. These instruments are frequently used to complement administrative education statistics and provide broader contextual information not available from available records.

Cross-sectional surveys are particularly valuable for describing access, participation and attainment in education, estimating socio‑economic, gender and regional inequalities, and identifying population groups facing barriers to educational opportunities. They support analysis of the relationship between education and social outcomes such as income levels, poverty, health and early childhood development. These insights play an important role in informing priorities for public investment, designing targeting mechanisms and guiding social policy decisions.

However, from a decision making standpoint, it is important to recognise the analytical limitations of this type of data. Because they observe different individuals at a single moment rather than following the same people over time, cross-sectional surveys cannot provide evidence on educational or labour market trajectories or on how outcomes evolve in response to policy reforms. They are not suitable for measuring the effects of programmes or interventions, as they cannot distinguish policy influence from pre‑existing differences between individuals, unobservable characteristics or broader macroeconomic conditions. As a result, analyses based solely on cross-sectional data can identify associations but cannot support credible causal conclusions about the impact of policies or programmes.

In practice, cross-sectional household and individual surveys offer significant value due to their representativeness, breadth of thematic coverage and periodic availability. They are particularly appropriate for diagnosing system-level inequalities, identifying priority groups and informing resource allocation. However, they are insufficient on their own for evaluating policy effectiveness, estimating programme impact or supporting high-stakes decisions such as scaling, redesigning or discontinuing interventions. For purposes requiring causal inference or trajectory analysis, they must be complemented by longitudinal, administrative or experimental evidence.

Labour Force Surveys

Labour Force Surveys (LFS) are nationally representative surveys implemented at regular intervals – typically quarterly or monthly – to monitor labour market conditions and employment dynamics within the working-age population. They collect detailed information on employment status, unemployment, job search activity, occupation, industry sector, type of employment contract, hours worked and forms of labour market participation. Because of their frequency and consistency, LFS datasets are a primary reference for understanding labour market performance and for designing employment and workforce‑related policies.

Examples include the European Union Labour Force Survey co‑ordinated by Eurostat, the Current Population Survey (CPS) in the United States, the  Labour Force Survey (LFS) in Canada, Enquête Emploi in France, ENOE in Mexico, GEIH (labour module) in Colombia, and similar instruments administered by national statistical offices across OECD and non-OECD countries. These surveys are widely used by labour ministries, public employment services and social security systems to track labour market dynamics, inform the design of employment and training programmes and identify groups requiring targeted support.

Labour Force Surveys are particularly valuable for understanding employment trajectories at the system level, examining unemployment patterns across demographic groups, analysing sectoral and regional labour demand, and identifying groups facing labour market barriers such as youth, women, migrants or older workers. They enable assessment of aggregate transitions between education and work, tracking indicators such as youth NEET rates (Not in Employment, Education or Training), inform the planning of training supply and support forecasting of workforce needs. Their ability to measure changes regularly makes LFS an important tool for policy response during economic shocks, recovery periods and structural labour market shifts.

However, there are important limitations to consider when using LFS data for policy decisions. Most Labour Force Surveys are cross-sectional, meaning they provide a snapshot of different individuals at a single point in time. Some operate as short rotating panels, but these typically follow respondents only for a limited number of waves and do not capture long-term changes. As a result, LFS datasets cannot observe individuals over extended periods or assess employment trajectories before and after a policy intervention, which limits their capacity to support analysis of long-term outcomes or causal effects.

Labour Force Surveys can describe correlations between qualifications and employment outcomes, but they cannot determine whether a particular programme, policy or reform caused changes in employment, productivity or earnings. Indicators captured in LFS are strongly influenced by broader macroeconomic fluctuations, demographic changes and institutional environments, which means that observed movements cannot be interpreted as policy effects without complementary evidence. In addition, LFS datasets generally lack granular information about programme characteristics, quality of training, employer behaviour or workplace practices, limiting their ability to explain why labour outcomes occur.

In practical terms, Labour Force Surveys are highly suitable for system-level labour analysis, labour market monitoring, workforce planning, identification of vulnerable groups and the design of employment and reskilling policies. They provide an essential evidence base for strategic decisions regarding labour supply, sectoral training priorities, and the organisation of public employment services. However, they are not sufficient for evaluating the effectiveness of specific education or employment programmes, nor for supporting high-stakes investment decisions, programme expansion or closure. For these purposes, LFS should be complemented by longitudinal data, administrative datasets, tracer studies or experimental and quasi‑experimental evaluations.

Administrative and assessment data

Administrative data

Administrative education data are collected routinely by education systems for operational and regulatory purposes. These datasets are typically derived from school, college or training institution records and include detailed information on enrolment, attendance, progression, completion, staffing, infrastructure and financing. Many countries consolidate these data into Education Management Information Systems (EMIS), which centralise information on students, institutions and programmes to support planning and system monitoring. Examples include Estonia’s EHIS, Norway’s VIGO and National Education Database, and Chile’s SIGE. Other countries have developed dedicated systems that support the connection, management and analysis of national education data. New Zealand’s National Student Index is one such example. These systems typically integrate enrolment, attendance, progression and institutional data, allowing administrations to monitor participation, track pathways and allocate resources more effectively.

Administrative datasets are particularly valuable for monitoring participation and progression within the education and training system, analysing dropout, repetition, transitions between levels and completion rates, and understanding the allocation of financial and human resources across institutions and regions. These data support evidence‑based planning, including forecasting demand for places in specific tracks or regions, adjusting funding formulas, and identifying institutions requiring targeted support. When integration across datasets is possible, administrative records can also enable analysis of medium- and long-term trajectories and provide the foundation for more robust evaluation designs.

However, administrative datasets also present important limitations that policymakers should consider. Because they are designed primarily for system management rather than analytical purposes, they often lack detailed information about students’ socio‑economic backgrounds, learning experiences or the quality of teaching and curriculum delivery. This limits the extent to which observed patterns can be interpreted or connected directly to policy actions. While administrative data enable strong descriptive analysis and trend monitoring, they do not provide causal evidence on the effects of reforms or programmes unless combined with appropriate evaluation designs or longitudinal linkages. Changes observed in administrative indicators may reflect multiple simultaneous influences – such as demographic shifts, economic conditions, classification changes or institutional incentives – rather than specific policy interventions.

Another consideration is data quality. As these datasets depend on reporting by institutions, accuracy may vary – incentive structures, workload pressures or inconsistent definitions can all introduce measurement discrepancies. Separately, the ability to link administrative education data with labour, tax or social protection systems depends on legal, technical and governance capacity, which may limit their potential for integrated analysis (see Section 5).

Standardised learning assessments and skills surveys

Standardised learning assessments and skills surveys are instruments designed to measure what learners and adults know and can do, using common frameworks, comparable scoring and consistent administration procedures. They enable meaningful comparisons across individuals, institutions, regions and countries, and support monitoring of performance over time. These assessments may focus on academic learning in school-aged populations or on foundational and workplace‑relevant skills in adults.

Examples of large‑scale international learning assessments include the OECD Programme for International Student Assessment (PISA), which evaluates 15‑year‑old students’ proficiency in reading, mathematics and science; TIMSS and PIRLS, which assess mathematics, science and reading respectively in school contexts; and national assessment systems used for certification or progression, such as the baccalauréat in France, ENEM in Brazil, the Saber assessments administered by ICFES in Colombia, or high-stakes end-of-school examinations in many OECD and non-OECD countries. For adults, the leading example is the OECD Programme for the International Assessment of Adult Competencies (PIAAC), which measures literacy, numeracy and problem-solving in technology-rich environments among individuals aged 16‑65. PIAAC also includes detailed background questionnaires on education, employment, job tasks and training participation. Countries also field national skills surveys to fill specific evidence gaps; Australia’s Foundation Skills Study, for instance, provides contemporary direct data on adult literacy and numeracy.

These assessments are particularly valuable for identifying disparities in learning and skill development across socio‑economic groups, genders, regions and school types, and for analysing how performance changes between cohorts or over time. They provide evidence for system-level policy prioritisation, support curriculum reform discussions, and offer comparative benchmarks that help countries understand their performance relative to peers. Skills surveys such as PIAAC also allow analysis of the relationship between skill levels and labour market outcomes, informing adult learning, reskilling and workforce development strategies.

However, it is essential to recognise the limitations of these tools when used to inform policy decisions. Standardised assessments focus primarily on the skills that can be measured reliably and comparably at scale. While some newer instruments include elements that capture broader competencies – such as socio‑emotional skills, collaborative problem-solving or aspects of creativity – the scope remains necessarily constrained. Interest in measuring socio‑emotional skills systematically is nonetheless growing, given their demonstrated links to educational retention, employability and well-being (OECD, 2025[10]). Many complex or context-dependent competencies, including professional practice or teamwork in real environments, are difficult to assess through standardised large‑scale testing.

While standardised learning assessments and skills surveys support analysis of associations between contextual factors and outcomes, they are not designed to determine whether a specific policy or programme caused the observed results. Many reforms operate simultaneously, and test score changes may reflect demographic shifts, implementation variation or external economic conditions rather than the direct effect of an intervention. Because most assessments are periodic rather than continuous, they are not suitable for short-term monitoring or rapid adjustment strategies.

Additionally, these instruments provide limited insight into how learning processes occur or why certain policies succeed or fail, since they do not collect detailed information on programme quality, pedagogical practices, resource allocation or institutional context. For these reasons, assessment results must be interpreted carefully, and policy conclusions should be supported with complementary evidence from administrative data, longitudinal datasets or rigorous evaluation methods.

Programme‑focussed data

Longitudinal and panel data

Longitudinal and panel data follow the same individuals, households or institutions over an extended period of time, collecting information at multiple points. Unlike cross-sectional surveys, which provide a single snapshot, longitudinal datasets enable analysis of trajectories, transitions and the evolution of outcomes before, during and after policies or interventions. These datasets may be formed through repeated surveys, linked administrative records, cohort studies or labour market panels, and can cover domains such as education, employment, income, social protection and health.

Prominent examples include the 1970 British Cohort Study (BCS70), which has followed over 17 000 individuals born in a single week of 1970 in England, Scotland and Wales. Initially focussed on perinatal health, it has evolved to provide rich evidence on health, education, labour market participation and socio‑economic outcomes across the life course. The British Household Panel Survey (BHPS) tracked approximately 10 000 households annually from 1991 to 2009 before being incorporated into Understanding Society, one of the world’s largest household longitudinal studies. Other examples include the German Socio-Economic Panel (SOEP), the US National Longitudinal Surveys (NLS) and registry-based longitudinal data systems such as Denmark’s DREAM database.

Longitudinal and panel data are particularly valuable for understanding how individuals progress through education and employment, how early disadvantages compound over time, and how participation in programmes is associated with medium- and long-term results. They support analysis of transitions – such as from school to work or from unemployment to stable employment – and allow examination of persistence, timing and turning points rather than static differences. These datasets provide a stronger basis for policy evaluation because they enable comparison of individuals before and after interventions and facilitate controlling for time‑invariant characteristics.

However, longitudinal studies also involve important limitations and trade‑offs. They typically have a narrower or more specific scope than survey-based national systems, as they often focus on particular cohorts (for example, youth, birth cohorts or specific programme participants). Their operational and follow-up costs are high, requiring long-term financial commitment, technical capacity, ethical frameworks and data governance. Attrition over time can reduce representativeness, particularly affecting disadvantaged groups, and fragmentation across systems may limit opportunities for interoperability and linkage. Importantly, although longitudinal data strengthen analytical potential, they do not generate causal evidence on their own. Credible causal attribution still requires a clearly defined evaluation design, suitable comparison groups and methods that can isolate policy effects from other factors. An additional consideration is timeliness: the long follow-up periods required to capture meaningful trajectories mean that by the time robust findings emerge, the skills needs of the economy may have evolved, potentially limiting the direct policy relevance of the results.

Tracer studies of graduates and programme participants

Tracer studies are surveys or follow-up studies conducted with individuals after they complete an educational programme, training intervention or social or employment support initiative. They typically collect information several months or years after exit, with the objective of understanding participants’ experiences, labour market integration, perceptions of programme quality and relevance, and short- to medium-term outcomes such as employment, earnings, job stability or further study. Tracer studies are commonly employed in vocational education and training (VET), higher education, active labour market programmes and social protection initiatives, particularly when institutions seek to understand participants’ early outcomes and experiences after completion.

Tracer studies are valuable because they provide programme‑specific evidence that is not available from national Labour Force Surveys or administrative systems. For example, they generate detailed information on participants’ perceptions of training quality, relevance of curricula to job requirements, and the extent to which skills acquired are used in the workplace. They support understanding of early labour market transitions, including entry pathways, employment barriers, and job quality. For providers, tracer studies inform programme improvement, career guidance services and employer engagement strategies.

However, important limitations must be recognised when using tracer study evidence for policy decisions. Tracer studies generally lack comparison groups, making it difficult to distinguish the effect of the programme from participants’ prior characteristics or labour market conditions. Without a counterfactual, observed employment or earnings outcomes cannot be interpreted as causal effects. Response rates may be uneven, and results may be biased if individuals with weaker outcomes opt out of follow-up, or if respondents differ systematically from non-respondents. Tracer studies also have a restricted scope, typically covering graduates of a single programme, institution or intervention rather than providing a system-wide view. Furthermore, they are costly and operationally demanding, as they require tracking participants over time, maintaining contact, ensuring data protection compliance and sometimes using mixed-mode follow-up strategies.

Evaluative and qualitative methods

Employer and enterprise surveys

Employer and enterprise surveys collect information directly from firms and industry representatives regarding labour demand, recruitment challenges, skills needs, training practices and the adoption of new technologies. They typically target business owners or human resource managers and may operate at national, regional or sectoral level. Because they capture employer perspectives, they offer insights into labour market requirements that cannot be observed through household or education data.

These surveys are especially valuable for identifying emerging skills shortages, estimating recruitment difficulties, and understanding how employers perceive the preparedness of graduates and workers. They help policymakers to align training supply with labour market demand, prioritise occupations with strong job prospects and adjust qualification frameworks or curricula in vocational and higher education. They can also provide information on firms’ own investment in training and internal workforce development.

However, several limitations are important to recognise when using employer survey evidence in policymaking. First, many employer surveys are sectoral, meaning that results are often not representative of the economy as a whole, and generalisation beyond the surveyed sector can lead to misleading conclusions. Second, methodologies, definitions and timing vary considerably across surveys and countries, which means results are rarely comparable across sectors or internationally. While international comparability may be less critical for a policymaker focussed on a specific domestic sector, inconsistent definitions can still limit the ability to benchmark performance or identify transferable practices. Third, responses often reflect employers’ perceptions, which may not coincide with objective skill requirements or performance indicators, and may be influenced by short-term pressures. Large firms tend to be over-represented, while micro‑enterprises and informal labour markets are frequently under-covered. Finally, employer surveys are generally not designed to establish whether specific programmes or reforms caused observed skills improvements, though they can capture employers’ perceptions of programme effectiveness, which offer a complementary – if subjective – evidence base.

Evaluative and qualitative methods

Data from experimental and pilot programmes

Experimental and pilot programme data are generated through interventions that are intentionally designed to test new policies, programmes or delivery models in a controlled or semi-controlled environment. These initiatives often include Randomised Controlled Trials (RCTs) or quasi‑experimental designs, in which participants are assigned to treatment and comparison groups. This allows the resulting data to support credible conclusions about cause – effect relationships.

Well-known international examples include the Jamaica Early Childhood Stimulation Study, one of the most influential RCTs in social policy. Beginning in the 1980s, it followed disadvantaged children who received early stimulation and parental coaching, comparing them with a control group. The study generated high-quality longitudinal data and demonstrated large and persistent gains in cognitive outcomes, schooling attainment and adult earnings, decades after the intervention (Gertler et al., 2021[30]; Gertler et al., 2014[31]). It remains a reference case worldwide for evidence‑based investment decisions in early childhood development. It is worth noting, however, that the intervention targeted stunted children in a disadvantaged community and included a nutritional component, which may account for a significant part of the observed gains. Caution is therefore warranted when generalising these results to broader early childhood programmes, particularly in high-income settings where stunting is less prevalent. Other examples include experimental evaluations of conditional cash transfer programmes and active labour market policies in Latin America (Escudero et al., 2017[32]).

These initiatives typically generate datasets that include detailed baseline characteristics, implementation records, and outcome measures such as learning gains, employment rates, job stability or earnings, often with follow-up data at multiple time points. The primary advantage of this type of data is its high internal validity, which allows policymakers to interpret observed differences as effects of the intervention rather than consequences of external factors. These datasets are also well suited for cost – effectiveness and cost – benefit analyses, supporting transparent and accountable resource allocation.

However, experimental and pilot programme data also have important limitations (Deaton and Cartwright, 2018[33]). Because these initiatives are usually implemented in specific locations, target groups or institutional contexts, the results often have limited external validity and may not generalise to national systems or broader populations. Sample sizes are typically smaller than those in national datasets, and follow-up periods may be short, which restricts the ability to capture long-term outcomes or spill-over effects. In addition, many RCTs and pilot interventions take place under controlled or highly supported conditions, with higher implementation quality, closer monitoring and stronger delivery standards than those that may be feasible at scale. As a result, effects observed in pilot settings may not fully reflect the outcomes that would occur once the intervention is expanded to more diverse institutions or delivered under routine system conditions. These studies are also costly, require advanced methodological and data management capacity, and depend on strong partnerships between institutions.

Qualitative data and case study research

Qualitative data and case study research involve systematic collection of non-numerical information through methods such as interviews, focus groups, field observations, classroom or workplace observations, and document analysis. Rather than aiming for statistical representativeness, qualitative approaches seek to understand mechanisms, behaviours, perceptions and implementation processes that are not visible in quantitative datasets. They provide depth and contextual understanding that support interpretation of quantitative results and inform policy and programme design.

Qualitative evidence is particularly valuable for exploring how and why policies and programmes operate as they do in real contexts. It can reveal perspectives from groups that may be under-represented or invisible in administrative datasets or survey samples, such as informal workers, learners outside the formal system or individuals with complex support needs. It can illuminate implementation barriers, institutional incentives, professional cultures and unintended consequences, providing insight into factors that influence programme success or failure. It also helps explain divergent results in quantitative analysis and generate hypotheses for further study.

However, qualitative evidence has natural limitations. Because it is not statistically representative, results cannot be generalised to the broader population without caution. Findings depend heavily on sampling strategies, researchers’ interpretation and participants’ willingness to share information. Qualitative approaches are not designed to measure programme impact or estimate effect sizes in the statistical sense. However, consistent patterns in participant-reported experiences can offer complementary evidence on mechanisms and perceived outcomes. They should not be used alone as a basis for high-stakes investment or reform decisions, but can strengthen the interpretation of quantitative findings. Collection and analysis can also be resource‑intensive when implemented at scale.

Emerging data sources

Big data, learning analytics and online job postings

Big data and learning analytics refer to large, high-frequency datasets generated through digital platforms and administrative technologies. These include data produced by online learning and training platforms, mobile applications and professional networking platforms. Because they capture real-time behavioural information rather than self-reported responses, they can provide highly detailed insight into participation patterns, engagement and performance.

A particularly important and rapidly growing source within this category is online job postings data, generated from digital job advertising platforms and aggregated labour market intelligence systems. These datasets collect information on vacancies, occupational profiles, required skills, credentials, technologies used and wage offers advertised by employers. OECD examples of research using these data, including analyses based on online job postings, include recent work by (Kuczera, 2025[34]; OECD, 2025[35]).

Big data, learning analytics and job postings are especially valuable for detecting emerging labour market trends, identifying skills in growing demand and informing curriculum modernisation and training priorities. Online postings provide near real-time insight into technology adoption, changing occupational profiles and local or sectoral shortages. Similarly, education platforms can track learner engagement, progression, completion and interaction with content, enabling adaptive learning models, early warning systems and targeted support.

However, these data have important limitations for policymaking. Online job postings capture only demand that is visible in digital markets, which means they overrepresent formal, large and high-skilled employers while underrepresenting micro‑enterprises, informal sectors and rural labour markets. Not all postings correspond to actual hiring intentions, and duplicated postings or automated listings can distort volume and trend interpretation. In addition, postings reflect employers’ expectations and preferences, which do not necessarily match actual job requirements or workplace practices. Learning analytics datasets contain only the behaviour of platform users and therefore exclude individuals without access or digital engagement, generating selection bias that can mask inequality. These datasets do not measure outcomes such as earnings, productivity or long-term employment and they cannot establish causality.

For these reasons, big data should be treated as a strictly complementary source: useful for horizon scanning and timely signals, but not a substitute for representative statistics or administrative records, and always with their limitations clearly documented.