Beyond the records: Data quality and COVID-19 vaccination progress in low- and middle-income countries
通过36个中低收入国家的电话调查和行政数据，发现调查得出的新冠疫苗覆盖率平均比官方数据高47%，主要源于抽样误差，纠正后仍有9个百分点的差距，表明行政记录存在缺陷。
Using phone surveys and administrative data in 36 LMICs, we find survey-based COVID-19 vaccine coverage to systematically exceed administrative figures by 47% on average. This discrepancy is strongest in Sub-Saharan Africa, raising questions about the role of data quality for our understanding of vaccination progress. We isolate the effect of sampling and measurement errors on the estimated vaccination rate in six survey experiments conducted in five Sub-Saharan African countries. Accounting for respondent selection effects reduces misalignment between data sources by 42% on average. Other commonly feared error sources including strategic misreporting, panel conditioning, or survey mode effects do not affect survey estimates. After adjusting for errors in the survey data, a substantial average gap of 9 percentage points remains with the official figures and seems to relate to flaws in administrative records that we document using data covering all 136 LMICs worldwide. Our results provide novel evidence on the size and sources of measurement error in modern data sources that have seen a surge in use by development economists over the last years. While our error-corrected estimates of vaccine coverage are still, on average, only a third as high in Sub-Saharan Africa as vaccination rates in high-income countries at the time, they imply that vaccination progress itself may have been quicker than what African countries were credited for in the public discourse. • Data quality issues impact our understanding of COVID-19 vaccination progress in a sample of 36 LMICs • Vaccine coverage estimated with surveys suggests much quicker progress than administrative records in Sub-Saharan Africa • Accounting for respondent selection effects in survey design reduces misalignment by 42% on average • Social desirability, panel conditioning, or survey mode effects do not affect estimates • Remaining differences seemingly relate to flaws in administrative records that we document with data for all 136 LMICs worldwide