Two new solutions to the zero weights problem

Abstract

Household panel surveys release inverse propensity non-response (IP-NR) weights that adjust selection weights for non-response. However, often some respondents lack selection weights, so cannot be assigned IP-NR weights. Previous solutions to this issue have limitations. We introduce two new procedures for weighting unweighted individuals. Both split the existing weights of individuals with unweighted individuals with similar characteristics. We evaluate procedure performance using data from the UKHLS COVID-19 Study, and find both perform well in increasing dataset sizes and reducing non-response bias and precision loss. We then discuss their broader use in survey design.