Donors Weren’t Spending a Lot on Evidence. It’s Now Going Off a Cliff.

Over the past two decades, the development community has made substantial progress in evidence generation. Donor agencies drove much of that progress by funding evaluations, research programmes, and the institutions that produce them. With recent cuts in aid, we are now at a critical moment where evidence generation is under threat. How much do donors actually spend on evidence generation? And, given the unevenness of cuts across development agencies, which donors spend the most?

Our findings suggest that evidence generation is in dire straits. Spending had been low: evidence generation accounted for less than 1 percent of annual disbursements from 2010 to 2025, even at the upper bound of our estimates. It is also concentrated: FCDO and USAID were the largest funders of evidence generation. Given that one agency is gone and the other is retreating, other donors will need to step up. And we need to do more to effectively use the evidence we do have.

Evidence spending is low, concentrated, and declining

We focus our analysis on the aid disbursements of nine of the largest development donors from 2010 to 2025: ADB, African Development Bank, FCDO, IADB, the Islamic Development Bank, the Ministry of Foreign Affairs of the Netherlands, Sida, USAID, and the World Bank. Using a large language model, we classify almost 80,000 development activities into rigorous evidence-generation, defined as any activity that includes research on the causal impact of a policy, programme, or intervention through rigorous counterfactual analysis. Our data come from the International Aid Transparency Initiative (IATI), which provides information on donor activities, including the activity title, description, and disbursement.

Figure 1 summarises our best estimate of evidence disbursement across donors. FCDO and USAID are the largest spenders on evidence generation. From 2010 to 2025, FCDO spent over $660 million on evidence generation—more than six times that of the third-highest spender, Sida. Across all donors, the magnitude of evidence spending is a very small share of total spending: evidence generation accounts for less than 0.6 percent of aggregate disbursements from 2010 to 2025.

Figure 1. Aggregate evidence disbursement by donor, from 2010 to 2025 ($USD)

Image

As seen in Figure 1, two key missing donors from our analysis are the Federal Ministry of Economic Cooperation and Development (BMZ) and Agence Française de Développement (AFD). While BMZ is one of the top ten largest donors in IATI, it reports disbursement data on only 12 percent of its activities. Similarly, AFD only began reporting its activities on IATI in 2018, midway through our analysis period. We therefore exclude both agencies from our analysis. For the World Bank, we include only non-trust-funded activities, to avoid double counting, so our results should be considered as what countries borrow to finance impact evaluation activities.

Over the fifteen-year period, spending on evidence has been low and falling. In Figure 2, we plot the estimated annual disbursement on evidence-generating activities over time. Estimated annual spending across all nine donors was less than $250 million, or less than 0.25 percent of annual disbursements. There is significant uncertainty around the point estimates, driven by both the data and the methodology (for more on this see our methodological note). Figure 3 shows the percent of disbursements devoted to rigorous evidence generation. Even if we were to account for uncertainty driven by classification error, the estimated evidence disbursement wouldn’t exceed 1 percent of aggregate spending.

Figure 2. Combined evidence disbursement across nine donors, by year, in $USD

Image

Figure 3. Combined evidence disbursement across nine donors, by year, as percent of total disbursement

Image

How we measured spending on evidence generation

These estimates rest on imperfect data and several key assumptions. Our dataset comes from the IATI, which provides information on donor activities, including the activity title, description, and disbursement. We used Claude Sonnet 4.6 to classify activities into rigorous evidence generation, defined as causal research involving rigorous counterfactual analysis. This includes randomised experiments and projects that use non-experimental methods, such as differences-in-differences, instrumental variables, or propensity score matching.

Of almost 80,000 activities from 2010–2025, the initial large language model (LLM) classification identifies 631 activities as including evidence generation. Donors vary in how they define activities and differ significantly in their reporting standards—both in the level of detail and in what they report. We therefore apply two corrections to the data that allow us to reach estimates of spending that can be compared across organisations and over time.

First, we adjust for potential misclassification errors. Activity classification relies on the accuracy of the LLM and the quality of the titles and descriptions reported on IATI. To account for this uncertainty, we estimate average error rates using a sample of 152 manually reviewed projects. We then use these to adjust for likely classification errors and to provide 95 percent confidence intervals on our estimates.

Second, we adjust for the disbursement value of evidence activities. Over half of evidence-generating activities are only sub-components of a larger development activity, as shown in Figure 4. For some organisations, like the World Bank, all evidence-generating activities are a component of a wider activity—there are no freestanding projects dedicated solely to evidence generation. The raw spending values for these projects would therefore significantly overestimate true spending on evidence generation. We adjust for this by valuing evidence-as-component projects at an upper bound of $1.4 million, the average cost of an independent impact evaluation (see methodology note for more details on how this is estimated). We apply these two corrections to estimate the size of disbursements on evidence generation from 2010 to 2025.

Figure 4. Initial LLM classification of evidence activities by donor and activity type, by $USD

Image

There are clear limitations to this approach, which we elaborate on in the methodology note. First, we are likely double-counting aggregate spending on evidence generation where there is substantial cross-organisational funding of evidence activities. For instance, organisations like USAID and FCDO have funded evidence-generation activities that were then conducted by the World Bank. This would likely appear as two separate activities on IATI—for example, an FCDO evidence generation project and components in World Bank projects. To help avoid this, we exclude World Bank trust-funded activities. Approximately 92 percent of these trust funds come from governments and development agencies; so we are likely to capture them in bilateral spending. Second, we are also likely underestimating evidence spending, as not all evidence related expenses are registered as development activities. This includes, for instance, the salaries of research staff within organisations that are not paid through projects reported in IATI. Lastly, we cannot ascribe a public or social value to evidence spending. The quality of evidence spending likely differs substantially across organisations. In related work analysing the top funders of development randomised controlled trials (RCTs), Leight finds differences in publication outcomes across donors.

Evidence generation at risk

The aid landscape ahead looks very different. FCDO, the largest spender on evidence, faces large funding cuts. And USAID, the second-largest donor, no longer exists. This is likely more damaging than just sheer spending levels alone: as Figure 1 shows, USAID and FCDO were significant funders of activities devoted solely to evidence generation. These activities often generated economies of scale —for example, learning what works for a particular issue and funding coordinated innovation and learning —as well as providing global public goods for governments and donor agencies.

While we hope the UK doesn’t cut this valuable contribution and we wait for a new US aid architecture to emerge, other donors will have to step up. Evidence on what works and what doesn’t is arguably more important than ever. In a world where development assistance as a whole is declining, more development activities will be self-financed. Knowledge generated by rigorous evaluation provides lessons not only for the country of implementation, but for the rest of the world as well. Subsidising this global public good of knowledge generation matters. Moreover, as we face constrained budgets—both domestically in low- and middle-income countries and from donors—the pressure to direct each dollar towards an effective intervention grows. The right evidence can help.

While spending over the last 15 years has been a low percentage of total development spending, it has generated a significant body of evidence. However, not all of it is relevant for decision-making, and what is relevant is often underused. Curating this evidence through tools such as smart buy lists and evidence advisory panels is one good investment to make use of development funds more efficiently. Another is better understanding how and when evidence is useful for decision-making, and using that knowledge to make more evidence-informed decisions and to shape what evidence gets produced. More on these to come.