Ever since the rise of randomized trials as a significant part of the economics literature, there have been recurring debates about the role of RCTs in the discipline. More specifically, many researchers debate whether the salience of RCTs incentivize a misallocation – or put more neutrally, a different allocation – of resources across methodologies, topics, and geographies. The hypothesized underrepresentation of development macroeconomics (or more broadly, research on economic growth) is often a focal point of these discussions.
But how common are RCTs? Is development today primarily trials of bednets and textbooks? With the benefit of new AI-empowered coding tools, these questions are very easy to answer.
Using AI to Code Up Development Papers 2021-2025
I recently generated a coding pipeline that seeks to identify the full set of development papers published in top general interest journals[i] as well as the Journal of Development Economics (generally considered to be the top-ranked field journal) between 2021 and 2025. (The exercise did not, however, include top field journals in macroeconomics or trade, where papers considered to be development macro might be published.)
Following the extraction of metadata (title, abstract, and keywords) for all papers, a Python script categorized them as development papers based on the JEL codes (in the O family), as well as a screen for any name of a low- or middle-income country or region in the title and abstract. Again, papers on development macro related topics that do not use O codes or highlight any reference may be more likely to be omitted using this strategy. This yielded a sample of 1600 papers; within this sample, the same script categorized papers based on whether or not they reported on a randomized trial (with any allusion to an RCT or experiment in the title or abstract).
Building on these basic characteristics of the sample, I then generated further scripts to categorize papers by topic; and by country of focus, if applicable. The analysis of topics included 16 different topic areas jointly identified by the analyst and the AI tool (one being an “other”, residual category): agriculture, behavioral and information, conflict / crime, education, environment, finance, firms, gender, health, infrastructure, labor, migration, political economy, social protection, trade / macro, and other (with “other” being the intended destination of, for example, papers focused on methods or measurement). Each paper is assigned a primary and secondary field, though in much of the analysis, I rank the two field assignments, if both are provided, as equally important (thus a paper that is identified as health and behavioral would count as 0.5 behavioral, 0.5 health; while a purely health paper with no identified secondary field counts as one). All analysis was primarily machine-driven with some human verification, and the full code pipeline and database is publicly available.
What did I find?
First, RCTs are not a majority of development papers. In the full sample, they constitute 25% of published papers, and this is relatively consistent over time. The highest percentage of RCTs is observed in the AEA-family journals, but it never exceeds 45%; the share varies in other top journals, and is 25% in the JDE. Compared to earlier estimates from McKenzie for papers published in 2025, the RCT share in top journals is virtually unchanged (29% vis-à-vis 31% in 2015) while the RCT percentage in three top general interest journals (ReStat, AEJ Applied, and EJ) has declined (28% in this sample vis-à-vis 44% in 2015). The most notable shift is at JDE, where the share of RCTs has roughly doubled, from 13% in 2015 to 25% in this sample.
What about the topics – do we see dominance of topics typically considered to be central in micro-development (education, health, gender)? This is also not the case – with the caveat, of course, that many papers cross multiple topics and the pipeline constrains the identified topics to be only two. Using the machine-generated classification, the most common topic is firms! The other four most common topics are labor, political economy, behavioral economics, and trade and macro. Health, education and agriculture follow as the next three most common topics. The composition of methodologies within topics also varies, but there is no topic in which RCTs constitute a clear majority of published papers (with only two topics, behavioral economics and finance), being approximately half RCTs.
Comparing across journal tiers (top five journals, other general interest journals, and the Journal of Development Economics), the mix of topics is relatively consistent. The only notable difference is that the prevalence of papers on macro / trade topics is notably higher in top five journals (15%) compared to other journals, particularly the JDE (7%), while agriculture has notably higher representation in the JDE (10%) compared to general-interest journals (less than 5%). This is consistent with the hypothesis that authors of macro and trade papers plausibly target different field journals.
As a final dimension, we can consider the distribution of country sites for empirical papers that are not cross-country. The analysis of country of focus drops papers that are purely theoretical and papers that use cross-country data (where cross-country is defined as more than five country sites). Relative to total population on a log scale, the representation of countries in the published literature seems broadly proportional, though with a number of countries in Latin America (Brazil, Mexico, and Colombia) as well as popular research sites (Kenya, Uganda, and Ghana) overrepresented, while Nigeria and Iran are notably underrepresented.
Relative to the total number of extremely poor people as measured in 2021, however, the pattern is more complex. The representation of India is still approximately proportional, with Latin American countries notably overrepresented, and a number of African countries (including Nigeria but also Mozambique, the Democratic Republic of Congo, Tanzania, South Africa, Malawi, Uganda, and Madagascar) notably underrepresented. China is dropped from the second graph; according to the World Bank, the number of extremely poor residents in 2021 was zero, and the graph is truncated at a minimum population of 100,000 extremely poor people.
To sum up: the development literature is incredibly diverse, and there is certainly no evidence it is dominated either by randomized controlled trials or by topics that are generally considered to be core micro-development topics (education, health, or agriculture). While some top-general interest journals disproportionately publish RCTs, this also varies, with macro and trade papers also highly represented in top five journals. Perhaps most important, sub-Saharan Africa is generally underrepresented relative to its size and relative to its concentration of poverty, particularly challenging or conflict-affected sites. In an environment of contracting resources and declining foreign aid, a renewed focus on these particularly underrepresented contexts may be important.
Jessica Leight is a Senior Research Fellow at the International Food Policy Research Institute.
[i] The journals included are American Economic Review, Quarterly Journal of Economics, Review of Economics and Statistics, Review of Economic Studies, Econometrica, American Economic Journal: Applied Economics, American Economic Journal: Economic Policy, Review of Economics and Statistics, Economic Journal, Journal of the European Economic Association, and American Economic Review: Insights.
Join the Conversation