Date of this Version
This dissertation focuses on understudied aspects of nonresponse in a context where limited information is available from refusals. In particular, this study examines social and psychological predictors of nonresponse in fast-paced face-to-face surveys; namely, election day surveys —popularly known as exit polls. Exit polls present unique challenges to study nonresponse since the population being sampled is fleeting and several conditions are beyond the researcher’s control.
If sample voters choose not participate, there is no practical way of contacting them to collect information in a timely manner. Using a proof-of-concept approach, this study explores a unique dataset that links information of respondents, nonrespondents, interviewer characteristics, as well as precinct-level information. Using this information, model-based plausible information is generated for nonrespondents (i.e., imputed data) to examine nonresponse dynamics. These data are then analyzed with multilevel regression methods. Nonresponse hypotheses are motivated by literature on cognitive abilities, cognition and social behavior.
Results from multiply imputed data and multilevel regression analyses are consistent with hypothesized relationships, suggesting that this approach may offer a way of studying nonresponse where limited information exists. Additionally, this dissertation explores sources of measurement error in exit polls. It examines whether the mechanisms likely to produce refusals are the same mechanisms likely introduce error once survey cooperation is established. A series of statistical interaction terms in OLS regressions —motivated by social interactions between interviewers and respondents— are used to explore hypothesized relationships. Overall, this research finds that cognitive mechanisms appear to account for voter nonresponse, whereas social desirability mechanisms seem to explain exit polling error.
Adviser: Allan L. McCutcheon