世界银行-提高纵向研究调查估计的质量_在LSMS面板调查中的应用（英）-2025.1_37页_917kb

Summarization

This technical memo poses an important question: how to improve the quality of survey estimates from longitudinal studies? We go beyond the current state and changes in human populations by focusing on the estimation phase. The challenge lies in maintaining accuracy as panel surveys extend over time, primarily due to sample erosion caused by deaths, movers, and new births, along with increasing sample fatigue contributing to measurement error.

We look at the case of panels with a rotating sample design, proposing a sophisticated change that blends tracking rules with the Generalized Weight Share Method (GWSM) and calibration estimators. Specifically, this approach addresses the biases in tracking by reflecting both the loss of representativeness and systematically following movers.

It further addresses complex household dynamics by refining the definition of households over time — from the traditional one-to-one continuity rules, to a many-to-many continuity rules (which capture one household potentially having multiple descendants over time, and vice versa). This flexibility ensures better alignment between estimated results and real-world population changes.

To implement this method, we introduce a base weight that recalibrates using auxiliary population data (like census figures and survey data) to account for new entrants and movers. Then, advanced calibration strategies—used in conjunction with standard statistical techniques—help mitigate non-response bias and maintain cross-sectional stability even as samples evolve.

In line with theoretical expectations and empirical tests from the Uganda National Panel Survey (UNPS) data, our results highlight the effectiveness of this method. The calibrated GWSM estimates produce more accurate individual-level statistics than traditional methods and show greater stability when cross-temporal samples are considered. Additional findings confirm the method’s robustness, particularly in addressing employment and poverty metrics where biases are significantly reduced.

Thus, while we report the work's methodological advancements and useful applications, we also emphasize the critical need for integrated datasets, refined tracking protocols, and periodic sample refreshes to sustain the longitudinal quality of estimates in ongoing field studies.