By Rozana Himaz, Dimitra Salmanidou and Saman Ghaffarian
When a tsunami or earthquake strikes, the immediate humanitarian response rightly dominates attention. But for researchers and policymakers trying to understand the economic aftermath and design better protection for future events, a less visible problem looms just as large: the data simply isn’t there.
To measure how a disaster affects household finances, we need to know how things looked before the event. Without that baseline, it is impossible to say how much worse off people became, or to rigorously evaluate whether aid and social protection reached those who needed it most. Standard imputation methods (ways for filling in missing data) assume some comparable data exists somewhere. But in most disaster-affected settings, especially in low- and middle-income countries, that pre-event data was never collected, or was destroyed in the disaster itself.
Our new paper tackles this fundamental measurement problem by combining machine learning with longitudinal household survey data. We hope that this approach can eventually scale toward the development of AI-driven systems capable of analysing diverse data sources for disaster welfare measurement.
A hidden cost of disasters: Out-of-pocket health spending
Natural hazards generate health costs that go well beyond the immediate emergency. Injuries, disease, displacement, and psychological trauma all drive demand for healthcare. At the same time, disasters destroy livelihoods and assets, squeezing households with rising costs and shrinking incomes. When health spending rises above 10% of a household’s income, economists define it as catastrophic, a threshold associated with falling into poverty and struggling to escape it.
Understanding how many households cross that threshold after a disaster, and by how much, matters for designing social protection systems. But answering that question requires a pre-disaster baseline, which is missing in most cases.
Using one disaster to understand another
Our approach uses two Indonesian household surveys to build a predictive model that can estimate what health spending would have been before a disaster struck.
Step 1: Training and comparing models
First, we train and compare models spanning four broad architecture types: linear, tree-based, neural network, and hybrid. We use data from the 2006 Yogyakarta earthquake, for which complete pre- and post-disaster survey data exist. The Indonesia Family Life Survey (IFLS) captured households both before and after the earthquake, allowing us to teach the model the relationship between household characteristics (income, wealth, family size, health status, education, location) and the share of income spent on health. We combine with this data information on hazard intensity.
Step 2: Applying the most robust model
Next, we apply the most robust model to a different setting: households surveyed after the devastating 2004 Indian Ocean tsunami that struck Aceh and North Sumatra. The chosen model is created using a combination of AI methods that work together to spot patterns and make predictions from data. The tsunami-specific survey (STAR) interviewed households months after the event but, like most post-disaster surveys, gathered little information about what life looked like beforehand. Our model fills that gap, generating plausible estimates of pre-tsunami health spending for each household.
This transfer across contexts (what we call transportability) is a core methodological challenge in itself. Can a model that was trained on earthquake-affected households in Java accurately predict health spending for tsunami-affected households in North Sumatra (Figure 1)?
We test this formally, examining how comparable the two survey populations are across measurable characteristics. The results show sufficient overlap to support model transfer for most households. However, there are meaningful contextual differences, particularly in the share of rural and farming households. This points to the longer-term ambition of this work: moving away from specialised hazard-specific assessments towards large-scale AI models that can be applied across many types of disasters and settings. By learning from a wider range of contexts, these models would rely less on any single imperfect comparison and, over time, produce more reliable predictions.
Figure 1: Geographic coverage of the two surveys. Tsunami-affected households are located across 13 districts in Aceh and North Sumatra; earthquake-affected households are drawn from provinces across Java.
What we found: Aid helped but the picture is complex
Using the model’s predictions of pre-tsunami health share, we compare what happened after the disaster with what would likely have occurred without it.
- Before the tsunami, around 4.5% of households in the affected areas were spending catastrophic shares of their income on health. Without any aid, that figure would have risen to 29.4%. Direct humanitarian assistance to households brought the share down to around 18%.
- Among the poorest fifth of households, nearly a third were already above the catastrophic threshold before the disaster. Post-tsunami, that rose to 46% without aid. Among the wealthiest fifth, no households were catastrophically exposed pre-tsunami – though 18% crossed the threshold afterwards without aid.
- One finding requires careful interpretation: households in moderately damaged areas consistently showed higher health cost burdens than those in heavily damaged areas. This is likely not because moderate damage was worse for health, but because humanitarian aid was heavily concentrated in the most devastated zones, and because many of the most seriously injured in heavily damaged areas did not survive to be surveyed. These factors make the two groups difficult to compare directly and serve as a reminder of the challenges of using disaster data.
Why this matters beyond Indonesia
The measurement challenge we address here is not unique to health spending or to Indonesia. Across disaster-affected settings worldwide, researchers and policymakers routinely face the same issue of missing welfare outcomes.
In principle, the same approach could be applied to food security, educational spending, or labour market participation. It could be extended beyond high-impact, rare events to the more frequent, lower-impact disasters that affect millions of people each year, such as seasonal floods. The approach can also be extended to measure post-disaster outcomes.
The key constraint is not the method itself, but the availability of comparable datasets. Building and maintaining high-quality geo-coded longitudinal household surveys, and designing post-disaster surveys that collect retrospective information on key pre-event characteristics, is therefore essential for the future of disaster economics.
A first step towards scalable disaster welfare measurement
This paper is a proof of concept. It demonstrates that machine learning can enable credible welfare measurement in settings where conventional data collection is absent. But substantial work still remains.
Validation across more events, hazard types, and countries is needed before the approach can be considered general-purpose. The integration of additional data sources such as administrative records, satellite imagery and cross-sectional national surveys could progressively improve both accuracy and breadth of application. In the longer-term, frameworks similar to large-scale “foundation models” in artificial intelligence, trained on data from many disaster contexts, might support more robust and widely applicable predictions.
Summary
When disasters strike, key data on household welfare is often missing, making it hard to measure impacts or target aid. This blog outlines research using machine learning and Indonesian survey data to estimate missing pre-disaster health spending. A model trained on the 2006 Yogyakarta earthquake is tested and then applied to the 2004 Indian Ocean tsunami, where no baseline data exists. Despite a complex setting, it achieves over 70% accuracy. The results suggest that without aid, catastrophic health spending could have risen sharply. The study shows how AI can help fill critical data gaps and support better disaster response.