Impact of unit and item nonresponse on population total estimation in asymmetric business survey data

Abstract

Nonresponse is one of the main sources of nonsampling error affecting the accuracy of survey estimates, particularly in business surveys characterized by highly asymmetric data distributions. This study investigates the impact of unit and item nonresponse on the estimation of population totals using simulation experiments based on enterprise-level research and development (R&D) statistics. Several weighting adjustment methods for unit nonresponse and alternative imputation techniques for item nonresponse were evaluated under different levels and mechanisms of missingness. Estimator performance was assessed using relative bias, coefficient of variation, and relative root mean squared error. The results show that increasing nonresponse levels substantially reduce estimator accuracy, primarily due to increasing relative bias rather than estimator variability. The effectiveness of weighting adjustment methods depends strongly on the response mechanism: class-based response probability estimation performed best under missing completely at random conditions, while random forest–based estimation produced more accurate results when response behavior depended on enterprise size. The performance of imputation methods also varied across missingness mechanisms, with mean-of-donors hot-deck, nearest neighbour, and regression-based imputation performing best under different scenarios. The findings provide practical guidance for selecting appropriate nonresponse adjustment strategies in business surveys with strongly skewed variables.