How does the rich-poor learning gap vary across countries?

Most of us place great value on education as a potential tool for improving social outcomes: increasing earnings, raising productivity, empowering women, and improving health outcomes. But one of the great tragedies in development is that educational outcomes for the affluent exceed those for the poor—everywhere.

The “social gradient” in education is steep. Those with higher socio-economic status (SES) achieve much better learning outcomes than those with lower SES. In my country, South Africa, children in the wealthiest quintile of schools perform one-and-a-half standard deviations better than those in the poorest quintile in the SACMEQ Mathematics test. (SACMEQ is an international school testing program similar to PISA in OECD countries, PASEC in West Africa or SERCE in Latin America.) Conventionally, that is interpreted as about a three-year learning backlog for children in poor schools—already by grade 6!

Steep (and often convex) social gradients along with similarly steep and convex returns to education (as in many Latin American countries) dampen social mobility. Combine that with weak economic growth, and you have a poverty trap. Escaping from this trap must surely start with improving education. So let’s consider social gradients in education.

It is impractical in school surveys to ask children about household income, but they can provide information on household assets. Filmer and Pritchett have shown more than a decade ago how to estimate wealth effects without expenditure data or tears, by deriving an asset measure that does a pretty good job of distinguishing along the lines of socio-economic status.    

Once households are ranked by SES using the asset index and this information is linked to performance on a cognitive test, the social gradient emerges. But it is not easy to compare social gradients across countries. When constructed for each country separately, the latent asset-based SES measure gives different weights to various possessions. These weights are unique for each country. However, if a common SES measure is constructed across different countries that participated in the same survey, the same weights are given for the same possessions in different countries. But owning a car, a television set, or a dishwasher do not mean the same in terms of economic status of households in Ethiopia as in the United States. Does ownership of a bicycle convey the same information on underlying wealth of a household in Mozambique as it does in Germany? Or a radio? These differences are less when the comparison is between two countries that are more similar in economic structure or level of development, such as Uganda and Tanzania.

The problem increases when trying to compare countries that participated in different surveys that may not have contained questions about the same sets of assets. How should we go about comparing the social gradients in Kenya that participated in SACMEQ, and those in Peru that participated in SERCE? Some of our current research tries to deal with this problem.

Our approach is deceptively simple:

  1. Rank children in each country using an asset-based SES variable derived for each country individually (i.e., separately for Botswana and for France, even though both participated in TIMSS).
  2. From household surveys for each country, children of similar age who are attending school are ranked by the per capita income or expenditure of their household, converted into purchasing power parity (PPP) dollars. (It would have been easier if all children were at school, but this is still far from common in most African countries.)
  3. Then it is simply a matter of matching: A pupil in the x-th percentile of the distribution of SES in the educational evaluation is allocated the per capita income of a child at the x-th percentile of the income distribution of those children of that age group who are in school. Essentially, we convert each country’s SES measure to a per capita income or consumption metric.
  4. Combining this with data from studies that convert international performance levels across surveys allows us to arrive at social gradients with a common scale on both axes.

Thus far our research has extracted results for grade 6 Mathematics students from two testing systems, SACMEQ and SERCE, for countries where survey data were available. A prominent provisional result is the remarkably good performance of Kenya. Even though it is a low-income country with limited educational resources, it not only outperforms much wealthier countries such as Uruguay, Costa Rica, and South Africa for given levels of per capita consumption, but also manages to provide more students with sufficient mathematical skills than these middle-income countries.

Furthermore, among poor children (under the $2 per person per day poverty line), Kenya’s educational performance outshines that of any other country we looked at, with 24 percent of poor students being mathematically competent or mathematically skilled, compared to 21 percent in Mexico and 16 percent in Peru, some of the better performers. Only 10 percent of poor students in Brazil achieve such competence levels. The situation is worse in Uganda (6 percent), Mozambique (5 percent), South Africa (2 percent), and Malawi (1 percent).

Another interesting result is the difference in the quality of educational outcomes between two very poor countries, Mozambique and Malawi. Mozambique has many more children living in poverty, but half of Mozambique’s poor children are proficient in numeracy, whereas few poor Malawian children are. In fact, even affluent students in Malawi perform poorly; the social gradient is remarkably flat. (This is true also for other social outcomes in Malawi too. Child nutrition, for instance, hardly improves with income. Not even the wealthy seem to benefit from quality public services in that country.)

Comparing some of these results again highlights weak service delivery in many developing countries. Even where resources may be similar, social gradients are steep in some, indicating much worse educational outcomes for the poor. And public resources are often extremely poorly converted into learning. South Africa spends multiples of resources per schoolchild as Malawi, and more parents are educated, yet performances on tests of poor South African children are very similar to those in Malawi. The differential ability of schools and school systems to convert resources into learning outcomes remains a major impediment to improving educational outcomes, and indeed life chances, for the poor.