1. Introduction: Poor learning costs the world $3 trillion per year

An investment in knowledge pays the best interest

– Benjamin Franklin in Way to Wealth (1758)

While he is perhaps best known for making the connection between lightning and electricity, for 25 years Benjamin Franklin also published a successful magazine that was focused on financial and life advice for the “common people,” as he put it. Called the Poor Richard’s Almanac and summarized in a 1758 essay entitled Way to Wealth, Franklin provided aphorisms and quotes that arguably are still relevant today. The quote above claims that investing in knowledge generates the greatest returns. Was Benjamin Franklin right? In this Best Investment Paper, we make the case that “Yes!” Franklin was right although with important caveats: the investments have to work, and they have to be backed by evidence that demonstrates their cost-effectiveness, characteristics that, unfortunately, are not always apparent in education policy. However, when they exist, investments in learning generate some of the highest returns across all sectors of development.

Until the 19th century, formal education was primarily for society’s elite, such as royalty, aristocracy, and leaders of religious institutions. Formal education as we know it today began expanding in Central and Eastern Europe during the 19th century (Lee & Lee, Reference Lee and Lee2016). Over the course of the Industrial Revolution and beyond, the share of people above age 15 with at least a basic education rose steadily from less than 20 % in 1800 to above 80 % by 2016 (see Figure 1).

Figure 1.

Share of the World Population older than 15 years with at least a basic education. Source: OurWorldInData. Global Education (OECD IIASA, 2016).

Today, efforts to formally educate populations have spread to every part of the globe. The latter half of the 20th century and the Millennium Development Goal (MDG) era saw a great push to boost primary school enrolment in low- and middle-income countries. In 1970, across developing nations, gross enrolment was 87 % (and only 51 % in low-income countries (LICs)), and now this is over 100 % with gender parity in schooling (World Bank, 2021). As a development initiative, much hope was laid at the feet of education – specifically enrolment, with the assumption that children would automatically learn – to boost the prospects of developing countries and help them grow through a better-educated labor force.

Unfortunately, we have since come to realize that being in school does not guarantee learning. While systematic time-series data on the main SDG goals are sparse, a recent estimate suggests only 1 out of 10 children in LICs are able to read by age 10,Footnote ¹ despite supposedly having attended school for 2–4 years (Azevedo et al., Reference Azevedo, Goldemberg, Montoya, Nayar, Rogers, Saavedra and Stacy2021). In comparison, 9 out of 10 children in high-income countries are able to read by age 10. A different but related measure of learning, the Harmonized Learning Outcomes (HLO), shows a slow improvement in primary school learning from 2000 to 2015 for most developing regions of the world, but the improvement is vastly insufficient to reach high-performing education systems by 2030 or even 2050 (Angrist et al., Reference Angrist, Djankov, Goldberg and Patrinos2020).

Lack of learning has profound implications for individual productivity and country-wide economic growth as well as development more broadly. Using the relationship that 1 standard deviationFootnote ² (SD) improvement in test scores – a common measure of learning – is correlated with a 20 % increase in income (Angrist et al., Reference Angrist, Djankov, Goldberg and Patrinos2020), exposure to an intervention that boosts test scores by 0.25 SD (a high but not extreme result) could increase an individual’s future earnings by 5 %. At a country level, a recent study in Nature indicates 20–50 % of income differences between nations are associated with differences in learning levels (Angrist et al., Reference Angrist, Djankov, Goldberg and Patrinos2020).

How much does poor learning in primary school cost the world every year? The loss can be approximated by noting that high-performing economies, such as the UK, may generate learning equivalent to 1.2 SD of test score improvements per year (Angrist et al., Reference Angrist, Djankov, Goldberg and Patrinos2020) while Evans and Yuan (Reference Evans and Yuan2019) show that each year of schooling in LICs or lower-middle-income countries (LMCs) generates 0.15 SD of test score improvements. A top school system such as the UK generates eight times as much learning as the typical LIC or LMC system.

If this gap – 1.05 SD – could be eliminated in LICs and LMCs for one year, it would boost the future incomes of school-age children by 21 %. Given a present value of lifetime income of $30,000 at an 8 % discount rate and 467 million children enrolled in primary school in these countries, the loss is equal to roughly $3 trillion per year. Essentially, if all LIC and LMC education systems could rise to the level of the UK then the annual benefit would be equivalent to ~40 % of combined LIC and LMCsʼ GDPs. Another way to put this is for an average country classified as LIC or LMC, a completely successful education reform could increase that country’s wealth at least 40 % before accounting for any spillover or general equilibrium effects. There is, of course, substantial uncertainty in this estimate. For example, the link between learning and incomes has only been empirically established cross-sectionally (and not more robustly, for example, via a randomized control trial with a long-term follow-up) with a relatively wide range around the strength of the relationship (Hanushek & Zhang, Reference Hanushek and Zhang2015; Evans & Yuan, Reference Evans and Yuan2019). Nevertheless, this calculation at least provides a plausible order-of-magnitude figure of the income loss from suboptimal learning.

This article sets out the economic case for two highly cost-effective interventions that can partially address this gap:

1. (i) Structured pedagogy, defined as a coherent package of inputs, typically includes textbooks, lesson plans, teacher training, and ongoing coaching of teachers that work together to improve in-class teaching, and

2. (ii) Teaching according to learning levels rather than age or grade (with and without technology).

Overall, the focus of this analysis is on primary education rather than secondary or tertiary levels of education. This is motivated by the fact that returns to education are generally higher at younger ages (Heckman, Reference Heckman2006) and that the unit cost of interventions in primary schools is lower due to higher levels of enrolment. Much of the evidence on interventions to improve learning comes from primary school settings.Footnote ³

The remainder of this article is structured as follows: In the next section, recent experience in the education sector in terms of both learning outcomes and spending is charted out and demonstrates the weak relationship between the two. What the world promised for the SDGs and an estimate of this cost based on historically ineffective approaches is then described. The conclusion is that meeting all the SDGs for education is unfeasible and expensive. In the next section, those interventions that are likely to be most cost-effective are discussed, and using existing reviews, 150 interventions are whittled down to the two most efficient examined in this study. A cost-benefit analysis of these two interventions is then discussed and finalized with a discussion.

2. The MDG era: Substantial success in boosting access but limited success in improving learning

The MDGs focused on increasing access to primary education, particularly for girls. These aims were achieved to a great degree with primary gross enrollment across the developing world above 100 % in 2015 (World Bank, 2021). The most substantial gains occurred in LICs, which went from 72 to 99 % enrollment. Gender parity in primary schooling was also achieved in 2015 (World Bank, 2021).

The results for learning were less impressive. The Harmonized Learning Outcomes (HLO) database provides information on how countries fared on various standardized tests across time and represents the most comprehensive dataset on learning outcomes globally (Angrist et al., Reference Angrist, Djankov, Goldberg and Patrinos2020). Focusing on various geographical regions, the results show that in much of the developing world, test scores did not improve substantially between 2000 and 2015 after adjusting for enrolment (see Figure 2). For example, in sub-Saharan Africa, the region with the greatest learning poverty, HLO scores only increased by 1.3 points per year. To put this figure in perspective, it would take another 159 years of improvement at the same pace to reach the level North America (539) had in 2015. Put differently, an average primary school student in sub-Saharan Africa would rank in the bottom 0.6 percentile of all primary students of the same grade in North America.Footnote ⁴

Figure 2.

Harmonized Learning Outcomes by region, 2000–2015 (primary school). Source: Angrist et al. (Reference Angrist, Djankov, Goldberg and Patrinos2020).

This lack of progress in learning outcomes (conditional on enrolment) is concerning given the huge resources devoted to improving primary grade education quality in LICs and LMCs across the same period. At the start of the MDG era, we estimated that LICs and LMCs spent $55 billion to keep 324 million children in primary school, at a cost of $170 per student (2020 USD). By 2015, total government spending per student, excluding costs associated with new enrolments, had increased to $307 (2020 USD) or an 80 % increase in spending (see Figure 3).Footnote ⁵ This is based on findings from two reports, the International Commission on Financing Global Education Opportunity (2016) and Al-Samarrai et al. (Reference Al-Samarrai, Cerdan-Infantes and Lehe2019).Footnote ⁶

Figure 3.

Increase in government spending per primary student across MDG era, LICs and LMCs (Breakdown by LICs and LMCs: LICs spent $39 per child on 42 million enrolments in 1998. By 2015, they spent an extra $30 per child more on 101 million enrolments. LMCs spent $190 per child on 283 million enrolments. By 2015, they spent an extra $165 more on 368 million enrolments). Source: Authors’ calculations.

These large increases in spending have not coincided with increased learning.Footnote ⁷ To understand why these efforts had limited success, it is necessary to consider how this money was spent. Here, data challenges are severe. There is no cross-country database on how education ministries allocate their budgets. Nevertheless, the few pieces of evidence paint a picture that suggests that for many countries, the basic and untested proscription was (and likely remains): more inputs = better education.

Prominent country examples of the fallacy that more inputs equal better education come from India and Indonesia, which together enroll a third of all primary school children in LICs and LMCs (UNESCO-UIS, 2021). In India, the Right to Education Act (2009) was heralded as a ground-breaking piece of legislation to address education outcomes in the country. The Act mandates a series of inputs with respect to the availability of schooling, teachers, remedial education and more while being silent on actual outcomes that need to be delivered by one of the world’s largest education systems. Inputs to schooling increased (Muralidharan, Reference Muralidharan2013), with the largest share of funding going toward teachers, both hiring more teachers to reduce pupil-teacher ratios and increasing teacher wages in real terms (Muralidharan, Reference Muralidharan2013; Bordoloi et al., Reference Bordoloi, Pandey, Irava and Junnarkar2020). However, there has been little to no improvement in learning outcomes. In fact, from 2006 to 2014, the share of children in third grade who could read a first grade text fell from 48 to 40 % (ASER, 2015).

Another piece of education reform occurred in Indonesia in 2005. This reform, called the Teacher Law, required teachers to pass a nonstringent certification process that then provided them with double their base salary (Ree et al., Reference Ree, Muralidharan, Pradhan and Rogers2018). Because of this reform and the fulfillment in 2009 of a constitutional commitment to spend 20 % of the national budget on education, the education budget in Indonesia doubled in real terms between 2000 and 2012, with two-thirds of the money being used on teachers (World Bank, 2013). The increase in salary had no discernible impact on children’s learning as measured by test scores (Ree et al., Reference Ree, Muralidharan, Pradhan and Rogers2018).

This brief synopsis has focused on teachers, but similar statements can be made about other inputs provided in isolation such as teacher training unconnected to classroom pedagogy and more textbooks and laptops (World Bank, 2020). Together, these provide insight as to why the challenge of learning has not yet been addressed. For LICs, the focus so far has been on access: 70 % of extra spending was directed toward this aim (Al-Samarrai et al., Reference Al-Samarrai, Cerdan-Infantes and Lehe2019). For lower- and upper-middle-income countries the main problem has been inefficient spending, with an outsized focus on more inputs. This is not entirely the fault of policymakers. Scientific understanding of “what works” in education is relatively new, with education randomized control trials proliferating only from roughly 2005 (Connolly et al., Reference Connolly, Keenan and Urbanska2018).Footnote ⁸ This work has questioned the efficiency of some norms in education such as organizing by age and providing more inputs in isolation as effective learning strategies. As we will demonstrate, inputs are important, but they cannot be provided in isolation. Instead, they need to be provided within the context of an intervention or program that demonstrates with evidence how the inputs together improve learning.Footnote ⁹

The SDGs cover multiple education aims – many cannot be implemented effectively, and if we continue with existing approaches, boosting learning will be costly to execute and is unlikely to succeed.

The SDGs include a myriad of education aims, including:

1. (i) Universal primary and secondary education leading to relevant learning outcomes.

2. (ii) Early childhood education and preprimary education.

3. (iii) Vocational, technical, and university education.

4. (iv) Relevant skills for work.

5. (v) Equal access to women and the vulnerable, including persons with disabilities, indigenous peoples, and children in vulnerable situations.

6. (vi) Universal youth literacy.

7. (vii) Education for sustainable development and global citizenship.

8. (viii) Education facilities that are child, disability, and gender-sensitive.

9. (ix) Substantial expansion globally of the number of scholarships available to developing countries.

10. (x) Substantial increase of the supply of qualified teachers, including international cooperation for teacher training in developing countries.

In the lead up to the ratification of the SDGs, Copenhagen Consensus conducted a rapid analysis of the preliminary goals and targets of the Open Working Group, the party responsible for crafting the SDGs. In this analysis, most of the proposed interventions were considered only fair or poor in terms of effectiveness (Lomborg, Reference Lomborg2015). Since then, a substantial body of evidence has been generated showing how certain interventions can increase learning at low cost (Angrist et al., Reference Angrist, Evans, Filmer, Glennerster, Rogers and Sabarwal2020; World Bank, 2020), motivating the focus on learning outcomes in this article.

Despite the many different goals for education, it is likely that given existing trends, the vast number of extra resources for education in the SDG era will be spent on more inputs in isolation. For example, the International Commission on Financing Global Education Opportunity (2016) provides one estimate of the extra funding required and mostly allocates this to “business-as-usual approaches”: higher salaries for teachers, lower pupil-teacher ratios, and expanding inputs, which given the evidence, are unlikely to make much of a difference to learning levels. Specifically, they suggest that by 2030, an extra $116, $246, and $1035 more is required per child for LIC, LMCs, and upper-middle-income countries, respectively. These represent, on average, an 83 % increase in spending on 2015 levels and based on current enrollment, would require an additional $330 billion per year by 2030. In the next section, evidence is provided on how the global community can increase learning levels at a much lower cost, requiring only $7.6 billion more per year above existing spending per child.

3. An overview of interventions to address learning: Which are the most cost-effective?

Doing a complete overview of the effectiveness of education initiatives is a challenging task given the proliferation of studies over the last 10 years. As a starting point, we take the cost-effectiveness review of 150 interventions in global education (Angrist et al., Reference Angrist, Evans, Filmer, Glennerster, Rogers and Sabarwal2020).Footnote ¹⁰ This was a background article for the Global Education Evidence Advisory Panel, an expert group convened to recommend interventions to improve learning in developing countries (World Bank, 2020). The panel, consisting of leading academics and practitioners in education that included Nobel Laureate Abhijit Banerjee, Yidan Prize winner Rukmini Banerji, and former Peruvian Minister of Education, Jaime Saavedra, categorized various interventions into four groupings: great buys, good buys, promising but low evidence, and bad buys. We also refer to previous cost-benefit analyses conducted by the Copenhagen Consensus Center as part of its global and country-level prioritization exercises.

With this in mind, the first takeaway from the review by Angrist et al. (Reference Angrist, Evans, Filmer, Glennerster, Rogers and Sabarwal2020) is that nearly half of the interventions show no effectiveness whatsoever. These interventions focus mostly on providing inputs alone, cash transfers,Footnote ¹¹ and generic teacher training. This finding broadly reinforces the results from Al-Samarrai et al. (Reference Al-Samarrai, Cerdan-Infantes and Lehe2019). The Panel categorized these interventions as “bad buys.” One clear recommendation is for governments to not adopt these ineffective policies and redirect funding toward more effective policies if their focus is on improving learning.

The Global Education Evidence Advisory Panel recommended only one intervention as a “great buy” based on the benefits, costs, and quality of education, specifically context-relevant information that shifts people’s beliefs about the benefits of education or the quality of schooling being provided.Footnote ¹² An example of this was the provision of basic statistical information to parents about the wage impacts of more schooling in Madagascar (Nguyen, Reference Nguyen2008).

The Panel also recommended the following interventions as “good buys”: structured lesson plans with linked materials and ongoing monitoring, coaching, and training of teachers (structured pedagogy); target teaching instruction by learning level, not grade (in or out of school); reduced travel times to schools (through school construction or bicycle transfers); giving merit-based scholarships to disadvantaged children and youth; using software that adapts to the learning level of the child (where hardware is already in schools) and preprimary education (ages 3–5).

Interventions categorized as “promising but low evidence” were the following: early childhood stimulation programs (for ages 0–2) that targeted parents, teacher accountability and incentive reforms, and community involvement in school management. Although there is less evidence for these educational interventions, preliminary findings showed they are highly cost-effective.

Following the recommendations of the Panel, we screened out the “bad buys” and the “promising but low evidence buys” from further consideration. We also screened out preprimary education because a recent review of BCRs for expanding preprimary learning indicates BCRs of 1.5–7.8 at discount rates of 5 % or higher (Horton & Black, Reference Horton, Black, Bundy, de Silva, Horton, Jamison and Patton2017). While respectable, the BCRs for these interventions are substantially lower than the BCRs of the interventions recommended in this article.Footnote ¹³

Last, we folded in “software that adapts to the learning level of the child” as a technology variant of “teaching instruction by learning level” as per the framing of Angrist et al. (Reference Angrist, Evans, Filmer, Glennerster, Rogers and Sabarwal2020).Footnote ¹⁴

This leaves four interventions remaining from the Global Education Evidence Advisory Panel list of great and good buys:

1. (i) providing information about the benefits, costs, and quality of education,

2. (ii) structured pedagogy with ongoing monitoring and coaching of teachers,

3. (iii) teaching according to learning level (with and without technology), and

4. (iv) reducing travel times to schools as well as giving merit-based scholarships to disadvantaged children and youth.

While these were all considered cost-effective by the Panel, in deciding their recommendations, the Panel only examined program costs per unit of learning gain (in this case, learning-adjusted years of schooling or LAYS). It appears they did not include all flow-on costs to education systems and students, particularly the costs associated with increased enrolment. These costs include the direct costs of schooling incurred by education systems that expand as enrolments increase. Alternatively, if education systems do not expand inputs (teachers and buildings) then the costs are borne by students who experience higher pupil-teacher ratios and potentially having to learn outdoors. For older children, there is also the opportunity cost of enrolment. We argue that conducting a cost-benefit analysis using this lens alters the calculus for interventions where the benefit is predominantly an increase in enrolment.

In this case, two interventions with a focus on increasing enrolment, these being reducing travel time to school and merit-based scholarships, will incur nontrivial flow-on costs beyond the program costs of providing these interventions. For example, a cost-benefit analysis on providing bicycles to girls in Rajasthan notes that extra schooling costs and opportunity cost of enrolment outweigh the cost of the bicycle by a factor of 2.3 (Mithal, Reference Mithal2018). This leads to a BCR of 6.6 at an 8 % discount rate, which is a good but not excellent return on investment.Footnote ¹⁵

A cost-benefit analysis of merit-based scholarships would represent, at best, the average return to schooling for the grade that the scholarship is being given. This is because the costs of such an intervention would be the direct and opportunity costs of schooling, while the benefits would be estimated by the increase in lifetime earnings from that year of schooling. This is the classic formulation of returns to education. The scholarship would be treated as a transfer. The BCRs from interventions designed to increase years of schooling are typically between 3 and 15 (see Appendix).

Last, we consider the overall evidence base and impact of including costs of increased enrolment for the Panel’s lone “great buy,” providing information on the benefits of schooling. Angrist et al. (Reference Angrist, Evans, Filmer, Glennerster, Rogers and Sabarwal2020), in their review, note only two studies with cost-effectiveness data for this intervention (Nguyen, Reference Nguyen2008; Jensen, Reference Jensen2010), each with very different cost-effectiveness outcomes. Nguyen (Reference Nguyen2008) generates a cost per learning-adjusted-year of schooling (LAYS) at 150 per $100 while Jensen (Reference Jensen2010) provides close to zero LAYS per $100 (see Figure 2 in Angrist et al., Reference Angrist, Evans, Filmer, Glennerster, Rogers and Sabarwal2020). In the course of reviewing an earlier draft of this article, one reviewer pointed out that Nguyen (Reference Nguyen2008) was not yet published and recommended we remove this article from the evidence base. That leaves just one paper with cost-effectiveness data.

However, additional papers have estimated impacts (though not with costs). As reported by Angrist et al. (Reference Angrist, Evans, Filmer, Glennerster, Rogers and Sabarwal2020), the average value of these is 0.08 LAYS. While moderate in size in line with typical effects in education around 0.1 SD (Evans & Yuan, Reference Evans and Yuan2022) it can be reasonably inferred to have a high return because the cost of providing information is very low.

This is where the importance of considering nonprogram costs becomes apparent. In the studies supporting this intervention, as cited by the Global Education Evidence Advisory Panel (World Bank, 2020), implementation of this intervention either increased enrolment by as much as 4.5 percentage points (Nguyen, Reference Nguyen2008; Jensen Reference Jensen2010; Andrabi et al., Reference Andrabi, Das and Khwaja2017; Gallego et al., Reference Gallego, Molina and Neilson2018) or compelled parents to choose more expensive schools (Neilson et al., Reference Neilson, Allende and Gallego2019). Given the modest impact and unaccounted costs, is the benefit-cost ratio (BCR) large enough to warrant inclusion in the recommended package? To assess this, we modeled the full costs and benefits of the intervention, including flow-on costs, estimating a BCR that was large (21) but significantly lower than the eventual BCR for the recommended package. Therefore, this intervention was also removed, although education ministries may want to consider implementing this as a next step after the recommended package since the BCR is still high.

That means that of 150 possible interventions, it is likely that these two are the very best:

1. (i) Structured pedagogy, defined as a coherent package of inputs, typically textbooks, lesson plans, and teacher training and coaching that work together to improve in-class teaching.

2. (ii) Teaching according to learning level rather than age or grade. This can be accomplished by technology that adapts to children’s learning level, or where this is not feasible (e.g., budget is unavailable for the sizeable upfront investment) by deploying “teaching-at-the-right level” as developed by the nongovernmental organization Pratham.Footnote ¹⁶

Of note is that there is evidence that these two interventions can be delivered at scale with government involvement (Stern et al., Reference Stern, Jukes, Piper, Stefano, Mejia, Dubeck, Carrol, Jordan, Gatuyu, Nduku, Punjabi, Van Keuren and Tufail2021), thus addressing a key challenge of the recommendations derived from randomized controlled trials of pilots and relatively small samples. These recommendations are therefore not merely academic in nature but are also supported by real-world success stories across Africa and South Asia.

Copenhagen Consensus has conducted multiple cost-benefit analyses on these and other interventions to improve learning as part of its global and country-level prioritization exercises. The results of these analyses are summarized in Table 1 and reinforce the broad findings of the Global Education Evidence Advisory Panel.

Table 1.

Findings from previous Copenhagen Consensus Studies

4. Cost-benefit analysis

4.1 General Parameters

The cost-benefit analysis considers a representative 8-year-old in a LIC or LMC. The student is exposed to the intervention and achieves increased learning, as measured in SD improvements in test scores, highlighted in recent reviews of investments to improve learning in lower-middle-income countries (LMCs; Evans & Yuan, Reference Evans and Yuan2019; Rodriguez-Segura, Reference Rodriguez-Segura2020). This increased learning leads to a 20 % boostFootnote ¹⁷ to counterfactual income per SD improvement following Angrist et al., (Reference Angrist, Djankov, Goldberg and Patrinos2020). Projected income is a function of GDP per capita that varies as countries become wealthier (see Appendix for estimation methodology). Individuals are assumed to work from age 18 to 64. The discount rate is 8 %. For the main results, we report findings from a rollout in a typical LIC or LMC. In the Appendix, we report additional scenarios including for upper-middle-income countries and using other discount rates.

There is the potential for substantial complexity in modeling the education pathways of an 8-year-old who receives the intervention versus one who does not. Key modeling considerations include the years of exposure to the intervention, the impact of learning on the probability of dropout in the years following the intervention, and the cumulative effect of these impacts (learning and years of schooling) on eventual school attainment.

While it abstracts from reality, we choose to model a simpler pathway that provides a cleaner cost-benefit test of an intervention’s impact. Specifically, we assume one year exposure to an intervention at 8-years-old and measure the gains against a counterfactual of average primary-level income. The alternative approach of modeling the full education pathway requires the use of additional assumptions (e.g., the probability of continuation being conditional on prior learning) that are not necessarily measured with high precision and for which effects would compound over time. This would generate complexity without necessarily adding any insight into these interventions’ effectiveness. Furthermore, many of the studies that provide effect sizes are only measured based on one year of exposure (or less) to an intervention, and long-term impacts are not well known.

Table 2 notes the most important parameters of the analysis and are explained below.

Table 2.

Key parameters