|
Abstract:
|
Survey nonresponse can be accounted for during weighting by using methods such as response propensity adjustments and calibration adjustments. These methods typically rely on the availability of adjustment variables related to an individual's response propensity and to key survey variables. However, complicated patterns of nonresponse can prevent the ability to control for all levels of relevant variables simultaneously, forcing compromises, such as using a simpler response propensity model or coarsening the level of data used in forming adjustment cells. These compromises may lessen the effectiveness of adjustments at reducing nonresponse bias. Further, sample size limitations could lead to increased design effects, given that small adjustment cells could lead to increased weight variation.
Using resampling methods, we explore the effects of sample size and nonresponse adjustment method on estimates in a post-election survey of overseas U.S. citizens following the 2014 General Election. In this survey, response rates were heavily impacted by voter participation history, state of registration, and country of residence. We found that holding weighting scheme constant, sample size reductions led to increased design effects that likely resulted from smaller adjustment cells. We also found that weighting schemes with lower complexity yielded a larger squared bias component of the mean squared error, particularly at larger sample sizes, although the simulated bias was fairly small and, therefore, not of practical importance.
|
