A Simulation Study of Alternative Weighting Class Adjustments for Nonresponse When Estimating a Population Mean from Complex Sample Survey Data

Results are presented from a simulation study of alternative weighting class adjustments for nonresponse when estimating a population mean from complex sample survey data, in an effort to extend the previous work of Little and Vartivarian (2003, 2005) to a complex sample survey setting involving stratified cluster sampling from a finite target population. A total of 30 simulations were performed, varying based on five different parameters: 1) the relationship of an auxiliary variable X available for respondents and nonrespondents with the survey variable of interest Y in the population of interest; 2) the relationship of the auxiliary variable X with the probability of unit nonresponse, P; 3) the use of base sampling weights according to a complex sample design when estimating response rates within weighting classes defined by X and/or the sampling strata (vs. no nonresponse adjustment at all); 4) the expected response rate for each sample across repeated sampling (75% or 25%); and 5) the relationship of the design strata with the probability of unit nonresponse, P. Each simulation examines the empirical bias and the empirical root mean squared error (RMSE) of a particular weighted estimator of the population mean for Y, with nonresponse adjustments to the base sampling weights computed in weighting classes defined by the auxiliary variable X and the sampling strata in selected simulations. Results from the simulations suggest that the use of weighted response rates within weighting classes defined by an auxiliary variable X and the sampling strata can be beneficial when working with survey data collected from stratified cluster samples, particularly when response rates are low and the auxiliary and stratum variables are correlated with both the survey variable of interest Y and response propensity.