In an effort to direct substantially more funding to the most cost-effective opportunities we can find to help people, GiveWell’s research is expanding in two directions at once: going broader, to cover more areas of global health and development, and going deeper, to improve the inputs that inform our decision-making.
This episode offers a behind-the-scenes look at how this expansion is playing out on GiveWell’s research team, as three researchers each describe a project they have taken on: reassessing how to prevent childhood pneumonia deaths, building a better estimate of how many children die of malaria, and developing an AI tool to check our work for errors.
In this episode, GiveWell co-founder and CEO Elie Hassenfeld speaks with Senior Program Officer Dilhan Perera, Senior Researcher James Watson, and Research Analyst Meghna Ray about the projects they’ve worked on and what it’s like to do research at GiveWell.
Elie, Dilhan, James, and Meghna discuss:
- Changing direction as we follow the evidence: Pneumonia causes around 10% to 15% of deaths among children under five globally. Our New Areas research subteam, which investigates cause areas that are newer to GiveWell, first examined community-based treatment options for reducing these deaths by treating children with amoxicillin, an inexpensive antibiotic. That approach initially looked promising, but conversations with researchers and implementers indicated that many pneumonia deaths today are caused by infections that amoxicillin doesn’t treat. We then shifted our attention toward another strategy for reducing pneumonia deaths: providing medical oxygen for severely ill children in hospitals. This costs more per patient but targets the children at greatest risk of dying and can also treat diseases beyond pneumonia. If expanding access to medical oxygen proves cost-effective, it could become a new way for GiveWell to help prevent these deaths.
- Improving our estimates for key cost-effectiveness inputs: Assessing the cost-effectiveness of malaria programs generally requires estimating the number of children who die from malaria in a given place, but that figure is surprisingly uncertain. In many places where we fund programs, cause-of-death data come mainly from interviews with families rather than from health system records. Because malaria symptoms are very similar to symptoms for many other diseases, knowing the symptoms a child was experiencing is often insufficient to determine whether malaria was the cause of death. To more accurately determine the share of deaths caused by malaria, we are developing an approach based on the share of school-aged children infected with malaria (which can be measured accurately using rapid fingerprick tests) and the overall child death rate (measured reliably using large household surveys). We then use the relationship between infection levels and deaths to estimate how many are attributable to malaria. Because this approach is based on data that we think can more reliably assess malaria-specific deaths, we expect it will be better at identifying differences in malaria deaths between regions, which could allow us to target malaria funding to the places where it will do the most good.
- Developing new tools to increase research capacity: Our cost-effectiveness spreadsheets are often very complex, and errors in their formulas or inputs can meaningfully change our estimate of a program’s cost-effectiveness, which is a key factor in our funding decisions. Research analysts have long vetted these analyses line by line to ensure their accuracy. We recently built an AI tool to assist in that process. Early versions reliably caught errors in formulas but often missed issues a human reviewer would catch, such as outdated data or inconsistent assumptions. To address this, we compiled errors analysts had found in past reviews and refined the tool until it caught more than 90% of them. The tool now serves as a first pass, followed by a research analyst reviewing its findings. As a result, vetting that took as long as a week now takes a few hours. Speeding up this check while maintaining accuracy frees researchers to spend more time on the substantive questions behind our grants.
This conversation explores aspects of our research process: building models, engaging with academic experts and implementing organizations, prioritizing the questions most likely to change our minds, and checking our work. This is part of what our researchers do, and we’re actively hiring for research roles so that we can keep identifying highly cost-effective programs and directing donors’ funds to the programs we think will do the most good.
Visit our All Grants Fund page to learn more about how you can support this work, and listen or subscribe to our podcast for our latest updates.
This episode was recorded on September 16, 2026, and represents our best understanding at that time.
Comments
I think it’s great that Give Well is rethinking its approach and both broadening and deepening its research on cause prioritisation and cost effectiveness – but I think the discussion here highlights how the path dependency baked into Give Well’s overall approach leads it to missed opportunities for more substantial impact.
I’m thinking particularly about the malaria example. It’s great to hear people talking more openly about how bad the data landscape is on mortality rates and causes of mortality in low income settings. As someone who has mostly done research in the humanitarian response space for the past decade, these are huge problems that don’t get enough attention.
However, the way the Give Well researcher seems to frame it, the problem he was tackling was to understand how many children were dying of malaria, in order to direct malaria resources more wisely.
But the only reason why we care about malaria in the first place is because we believed – on the basis of best available data at the time – that it was a primary cause of death for children under 5. If the data for that was unreliable, surely it suggests that cause prioritisation for malaria might be affected – in other words, the problem we started off trying to address was mortality rates for under 5s, and then focused on malaria prevention and treatment as one of the most cost effective pathways due to our belief about malaria’s contribution to global under 5 mortality rates (as well of course as other factors involving the relative low cost of malaria prevention and treatment solutions).
Perhaps Give Well has already addressed the question of comparing interventions that target all-cause mortality vs. cause specific mortality somewhere (for example, CHW delivery models vs. disease specific treatments). For me, the issues with malaria specific mortality data begs the question of why we are solving for a particular cause of death, if we care about excess mortality for under 5s and if the data on that particular cause of death is poor.
Surely if we have reason to suspect the data on malarial-related deaths, and we care about under 5 mortality, the higher impact approach would be to improve our understanding of all causes of under 5 mortality rates in low income settings, which could contribute to far better cause prioritisation in the short to medium term, and could also help identify better preventative care models overall. So, for instance, looking at point of care for children who die under the age of 5, understanding treatment protocols for a wider range of preventable causes, and assessing costs for these, would provide a much stronger data and evidence base than simply trying to solve for the question of which areas have higher rates of malaria-related deaths.
Thanks, Alice, for your comment. GiveWell aims to help save and improve lives as much as we can, regardless of cause.
While we have reason to question data on malaria-related deaths, our review of surveys, estimates from several different sources, and conversations with experts leads us to think that malaria is a major contributor to early childhood mortality. The alternative malaria mortality estimation approach that James describes in the episode, which suggests that about 40% more young children die from malaria than estimated by other data sources, provides some validation that GiveWell is correct to focus a portion of its attention on malaria prevention. In addition, relatively inexpensive anti-malaria interventions, like insecticide-treated nets and seasonal malaria chemoprevention, have substantial evidence of effectiveness and large funding gaps.
Because we continue to believe malaria is among the largest preventable causes of death with inexpensive and tractable prevention strategies, we want to invest in those strategies we believe are cost-effective and will have a substantial impact. We also recognize that some of the information guiding our decision-making is less robust than we’d like—we plan to continue stress-testing it and expect to use what we learn to improve future grantmaking decisions.
Our researchers are also doing work across a range of cause areas beyond malaria prevention and treatment. In addition to ongoing work to increase the number of children being vaccinated and funding programs to prevent and treat diarrhea, for example, we’re also learning about and addressing many other causes of child mortality. For example, as Dilhan describes in this podcast episode, we’re funding programs to increase access to medical oxygen. We’re also funding programs to support primary healthcare, learn more about whether supply chain digitization programs reduce stockouts and increase access to treatment, assess the effectiveness of alternative treatments for hospitalized children with severe acute malnutrition, and much more.
As part of this work, we are seeking to improve our understanding of the full spectrum of causes of under-five mortality to help us assess future opportunities to reduce child deaths. We recently made several related grants, including grants to produce state-level estimates of causes of death for children ages 0-5 in Nigeria from 2000-2024 and to collect nationally representative malaria biomarkers and under-5 verbal autopsy data in Côte d’Ivoire.