Bridging the survey gap: a call for innovative methods to predict poverty

Measuring poverty is important for development. If we don’t know who is living in poverty, we can’t design the right policies to help them or track whether those policies are effective.

Yet in too many countries, we lack timely, reliable, and comparable poverty data.

In some of the world’s most populous and vulnerable economies, comprehensive household surveys — the gold standard for poverty measurement —  are conducted only once every five or even ten years. In others, surveys may be delayed or disrupted by conflict, climate shocks, or political instability. Even when recent surveys exist, they often lack the detailed consumption data needed to estimate poverty.

The challenge of predicting poverty

To fill these gaps, researchers and practitioners can use different types of methods. One method that teams at the World Bank and elsewhere increasingly rely on is survey-to-survey imputation — a method that predicts poverty in recent surveys by modeling patterns from older ones.

This method has already been used to estimate poverty in countries like India and Nigeria, where up-to-date or comparable consumption data is missing. It has become an important tool to help monitor global poverty trends, support governments, and evaluate policy impacts over time.

But imputation is not foolproof.

As our recent research shows, even small errors in modeling assumptions—like changes in survey design or economic structure — can lead to large errors in predicted poverty rates. These risks are amplified when imputation is performed over time, in fast-changing contexts, or during times of crisis. Without rigorous testing, we may be misled by results that appear plausible — but are fundamentally wrong.

A competition to strengthen poverty prediction

That’s why the World Bank is launching a global modeling challenge, hosted on DrivenData, to gather insights from experts around the world to improve our tools for using survey-to-survey imputation to predict poverty.

This competition simulates a common challenge facing World Bank economists who monitor poverty:

• ·You’re given data from a previous year, including detailed information on household consumption.
• ·You’re asked to predict poverty in a new survey — one without consumption data.

As in real-world applications, participants will not receive immediate feedback on the accuracy of their predictions upon submission. However, unlike in real life—where the performance of your estimates may remain unknown — the true outcomes will be revealed after the final submission. You will be able to stress-test imputation methods under real-world constraints — changing survey populations, uncertain model validity, and limited feedback. Participants will build models that predict both household-level consumption and poverty rates across the population, using only the limited information available in actual surveys.

The results from this competition will help us answer key questions:

• What modeling strategies work best when survey data is limited?
• How do different techniques handle structural shifts in the economy?
• Can we reliably identify households living in poverty when consumption isn’t observed?

Why does it matter?

Timely poverty data is the yardstick by which we measure the impact of our work. We need to know not only who is living in poverty, but how poverty is changing in the face of climate shocks, demographic transitions, and instability.

The poverty estimates we produce shape real-world decisions — from national safety net programs and disaster relief to long-term development financing. Better models mean more responsive policies. More accurate estimates mean fewer people slip through the cracks.

Whether you’re a data scientist, economist, or simply passionate about better policies, we invite you to join us in our efforts to improve poverty measurement and help ensure that no one is left behind.

Want to learn more?

Our paper, Stress-Testing Survey-to-Survey Imputation: Understanding When Poverty Predictions Can Fail is a deep dive into when and why survey-to-survey methods work — and when they don’t. We encourage participants to read it and bring its insights into the challenge.

The Survey-to-Survey Imputation Challenge is now live.
Learn more about the challenge and join the effort at: https://www.drivendata.org/competitions/305/competition-worldbank-poverty/. Submissions are due by February 4, 2026.

As the competition unfolds, we’ll share updates on participation—highlighting the number of competitors, their geographic diversity, and their performance on a subset of the test data. However, please note that these interim results may not reflect final rankings, as they are based only on a portion of the evaluation dataset.