经济冲击下贫困监测调查：代理人和灵活非线性学习者的作用

11359 Survey-to-Survey Poverty Monitoringunder Economic Shocks The Role of Proxies and Flexible Nonlinear Learners Nobuo YoshidaYusaku KawashimaShinya Takamatsu Development EconomicsDevelopment Data GroupApril 2026 A verified reproducibility package for this paper isavailable athttp://reproducibility.worldbank.org,clickherefor direct access. Policy Research Working Paper11359 Abstract fast-changing proxy variables that track unobserved wel-fare changes are included, out-of-sample validity can berestored under well-defined conditions. Poverty estimation,however, requires a stronger condition—stability of the fullconditional distribution of welfare—beyond that neededfor mean welfare estimation. Monte Carlo simulations andevidence from Afghanistan, Uganda, and Rwanda supportthese predictions. The findings imply that improving ques-tionnaire content is more important than increasing modelcomplexity for timely poverty measurement in the face ofshocks. In many low-income countries, consumption surveys aretoo infrequent to track poverty during economic shocks.Survey-to-survey imputation can fill this gap, but its reli-ability depends on whether prediction models estimatedin one period remain valid in another. This paper developsa formal identification framework for the transportabil-ity of survey-to-survey estimators and shows that failuresduring crises arise primarily from missing shock-responsiveinformation rather than from insufficient model flexibility.When such information is omitted, both linear modelsand flexible learners yield biased poverty estimates. When The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about developmentissues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry thenames of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely thoseof the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank andits affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Survey-to-Survey Poverty Monitoring under Economic Shocks:The Role of Proxies and Flexible Nonlinear Learners Nobuo Yoshida, Yusaku Kawashima, and Shinya Takamatsu1 Keywords:survey-to-survey imputation; poverty monitoring; economic shocks; proxy variables; machinelearning; nonlinear models JEL codes: I32, C83, C53, O15 I.Introduction Household survey data—the cornerstone of official poverty measurement—are collected only sporadicallyin many low-income countries. Lanjouw and Yoshida (2021) document that nationally representativesurveys capable of producing consumption-based poverty headcounts appear, on average, once every sevenyears. As a result, when economic crises or climate-related shocks occur, no fresh survey is typically in thefield, leaving policy makers without timely information on poverty dynamics precisely when it is mostneeded. Survey-to-survey (S2S) imputation offers a practical response to this constraint. Instead of re-administeringa full consumption module—which can take one to two hours per household—S2S relies on a shortquestionnaire, typically consisting of 10–20 mostly binary items that can be collected quickly and at lowcost through phone or other remote modes. A model estimated on the most recent full survey maps theseresponses into predicted household welfare and poverty rates (Yoshida and Aaron 2023). The appeal ofS2S lies in its speed, scalability, and ability to provide near–real-time poverty estimates cost-effectively.2 The central challenge of S2S imputation, however, is out-of-sample validity. Large shocks may alter therelationship between welfare and its observable correlates, rendering models trained in calmer periodsunreliable. Corral et al. (2025) show that when such parameter instability occurs, S2S projections canbecome severely biased. Importantly, this problem is not tied to a particular estimator: any model trainedon pre-shock data will fail if the mapping between observables and welfare is no longer stable. Two broad responses to this challenge have emerged in the literature. One approach emphasizes functionalflexibility. If the conditional relationship between welfare and observables is nonlinear or complex,replacing linear projections with flexible nonlinear learning (FNL) methods—such as random forests,boosting, or neural networks—may improve predictive performance. A second approach focuses insteadon information content. Several empirical studies argue that breakdowns during crises arise becausestandard covariates are slow-moving and propose enriching S2S models with fast-changing poverty orwelfare correlates—such as short-term consumption indicators, labor disruptions, or subjective well-being—that respond directly to sh