A comparison of methods for combining surveys

Abstract

As more and more researchers start to study more complicated survey problems, a single survey might not be sufficient to meet the analytical needs. Therefore, combining multiple waves of repeated surveys to get larger and more diverse samples is useful in some situations. This thesis focus on comparing different methods of combining multiple waves of repeated surveys. A large simulation study where we investigate the properties of different methods for combining surveys is presented in this thesis. In the simulation, we create a series of populations that contain geographic information so the stratified two-stage cluster sampling method can be applied when selecting samples. Then, two methods (weight-adjust method and calibration weighting methods) of combining surveys are applied. The analytical results of simulated data show that the weight-adjust method can reduce the estimation bias (i.e. underestimating or overestimating) if overlapped observations can be identified and need to be removed in order to meet study needs when combining surveys. However, if the variables we are interested in change over time, the estimates produced by the weightadjust method may contain minor biases. Therefore, using the calibration method when auxiliary variables are available is recommended since it will help reduce the standard error and produce more accurate estimates. Especially when using re-calculated weights to estimate the variables that change over time, using the calibration method can reduce the bias.