Survey designs and compensation methods for nonresponse problems

• Main Author: Siripornpibul, Taweesak
• Language: en
• Published: University of Canterbury. Mathematics and Statistics 2011
• Online Access: http://hdl.handle.net/10092/5496

This thesis is a comparative study of the unit nonresponse problem. Simulation based on the summary report of the 1997 Thailand Industrial Survey is used to investigate efficient survey designs. Research questions are (1) Which design is best for conducting large-scale surveys such as the Thailand Establishment Survey with nonresponse problems? (2) Which compensation method can reduce the effects from nonresponse problems? This study begins with an extensive review which brings together the theory of the methods for compensating nonresponse. The simulation compares across all combination of survey designs and compensation methods, thus extending the work of other researchers. In the simulation study there are various sampling designs and compensation methods. Other factors that are considered are sample size, response rates, and types of response (dependent and independent). Nonrespondent subsampling is used to compensate unit nonresponse during the data collection phase. Nonrespondent subsampling methods used in this thesis are one- and two-subsampling schemes. Weighting adjustment procedures and imputation methods are used in the estimation phase. The response mechanisms for weighting adjustment procedures and imputation methods are a naive model and a random homogeneity (RHG) model. In addition a new weighting adjustment method called the bias-removal adjustment method is proposed. Methods for dealing with nonresponse are compared by using bias, variance and design effect. The main conclusion were, whenever possible, a complex survey design should be used, e.g. stratified or post-stratified sampling with unequal probability of selection. The best method to compensate for nonresponse is nonrespondent subsampling. If subsampling is too costly RHG model with weighting adjustment or with imputation is recommended. For example in weighting adjustment method, the population-based should be used in equal probability sampling with or without replacement. For unequal probability sampling, the sample-based methods should be used in sampling without replacement and the bias-removal method should be used in sampling without replacement. In imputation method, multiple imputation with regression or methods related with regression should be used combined with weighting adjustment procedures described above for each survey design. An algorithm for dealing with nonresponse is presented.