Minimising Total Survey Error Arising from Sampling and Non-Sampling Errors: A Design Based Approach

Abstract

This study was based on the total survey error paradigm in which the purpose was to examine a variety of sources of errors in a survey. The study sought to quantify the total survey error for estimating population total in two stage cluster sampling and empirically compute the mean square error estimate using data from an actual survey. Furthermore, the study analysed a method of handling non-response in complex survey design. The methodology of the study was centred on the decomposition of the mean square error into sampling error, refusal error, noninterview error and response error. These were further simplified to come with a total survey error expression. An empirical study with data based on waiting time to conception in women in Zimbabwe was conducted to estimate the population total survey error using this model. Data collection was done through face-to-face interviews in the selected clusters. The results shows that both unit and item non-response contributed significantly (90.8%) to the total mean square error. Imputation was done in order to reduce the error and the results show that imputation should be done within each cluster since the non-response depends on a cluster level random effect.