Measuring human capital using global learning data

We present a few descriptive trends in a first application of the database. The average learning outcomes for 164 countries from 2000 to 2017 is shown in Fig. 1. The global coverage of the database becomes immediately apparent and regions typically excluded from international tests such as PISA and TIMSS included in our database are clearly shown (Fig. 1). The database covers the vast majority of countries in sub-Saharan Africa, Latin America and the Caribbean, and South Asia— economies with considerable potential to close learning gaps.

A few trends emerge: high-income countries far outpace developing economies. Sub-Saharan African lags behind all regions besides South Asia, with countries such as India performing similar to lower-performing sub-Saharan African countries; within sub-Saharan Africa, a few countries such as Kenya and Tanzania lead, on par with many countries in Latin America. Within Latin America, a few countries such as Chile are on par with European countries. The Middle East performs similarly or worse than Latin America and many Asian countries outperform North American and European countries.

The expected years of schooling and HLO primary learning scores for the most recent year for which data are available are shown in Extended Data Fig. 1 to contrast the quantity and quality of the education system (detailed descriptions of the variables and analyses of the trends are provided in the Methods). The graph shown in Extended Data Fig. 1 shows that although many developing countries have achieved substantial levels of schooling (a typical education system is expected to deliver 10–12 years of schooling), they have not yet realized high rates of learning (advanced proficiency in international tests is centred around a score of 625). Two examples with high schooling but low learning are Brazil and Ghana. Brazil has 11.7 years of expected schooling, yet a learning score of just 426. Ghana has 11.6 years of expected schooling, yet a learning score of only 229.

We next explore the contrast between changes in schooling and changes in learning over time. We measure schooling using adjusted enrolment ratios. We compare this measure of schooling to our measure of learning in primary school for the years 2000–2015. We use data for this period as it has the highest overlap of schooling and learning measures. We restrict our comparison to countries with data points in at least two time periods for enrolment and learning in primary school to maximize comparability over the time period. We further condition on country-fixed effects using multivariate regression for each region (see Methods for details). This accounts for potential changes in the sample composition of countries with available data for each time period.

We observe a clear trend towards increased schooling, while learning progress appears to be limited in many cases. For example, in the Middle East and North Africa enrolment rates achieved a high of 99% by 2010, increasing from 95% in 2000. By contrast, learning levels stayed low and remained the same around an initial score of 380 from 2000 to 2015 in these regions. It is possible that in regions such as sub-Saharan Africa, as enrolment gets substantially higher and new, lower-performing students participate in learning assessments, average scores decrease owing to a selection effect. However, we observe slow learning progress even in regions in which enrolment levels are relatively constant and high, such as Latin America and the Caribbean, suggesting that there is more to these trends than selection. In Extended Data Fig. 2, we explicitly condition on enrolment and find nearly identical patterns. Moreover, a regression of primary learning outcomes on primary enrolment rates using a panel of countries between 2000 and 2015 with country-fixed effects yields a negative coefficient on enrolment rates of 0.247 with a P value of 0.673, further showing that higher levels of school enrolment has no statistically significant association with better learning outcomes (see Methods for details).

The patterns in Fig. 2 and Extended Data Fig. 2 could be interpreted as indicative of a plateau effect, as at higher learning levels—for example, in North America or Europe—obtaining further gains may be difficult. However, we also see a relatively flat line in cases in which baseline learning levels are low—that is, in Latin America and the Caribbean—which suggests that learning progress is slow regardless of initial learning conditions. Data availability for each country could in principle affect the patterns that we describe, but the robustness of the patterns to the inclusion of country-fixed effects as described above suggests that they are not driven by country composition. Country-specific trends are also illustrated in Extended Data Fig. 3.

Taken together, the data reveal that the rapid pace at which schooling improved across education systems, as well as the success in achieving high enrolment rates, have not been met with similarly rapid learning progress. This pattern has been referred to in the literature and by the international education community as ‘the learning crisis’. Using the HLO database, we demonstrate that it holds on a global scale.