Investigating the Bias Properties of Alternative Statistical Inference Methods in Mixed-Mode Surveys.

Abstract

Early in the history of survey research, mixed-mode surveys were proposed to decrease non-observational survey errors under certain survey budgets. The statistical inference in the earlier studies implicitly assumed ignorable mode effects; that is, responses to all survey modes generate values close to true values for all the members of the population. Inference in later mixed-mode survey designs, generally adopted the early assumption that mode effects could be ignored and did not challenge that assumption with any empirical work despite the developed theoretical frameworks. In practice, survey modes are not randomly assigned in mixed-mode surveys. This nonrandom assignment establishes a challenge to evaluate mode effects directly in mixed-mode surveys. This dissertation defines this nonrandom assignment as mode choice. Under the mode choice mechanism, an alternate method is proposed to evaluate and adjust for mode effects. In particular, the respondent data for a given mode and phase are used to create complete datasets for a given sample. Then, the complete datasets are used to compute mode-specific survey means that are then combined to produce one survey estimate. The mean estimates can be combined as (1) a simple average, (2) a minimum variance combination, and (3) a minimum mean square error combination. The last of these requires some measure of true values that are unaffected by mode effects. The dissertation includes conceptual work and empirical/simulation evaluation of inference methods. The conceptual work includes extension of a single survey mode statistical error model to a mixed-mode survey context. The bias properties of the standard method of estimation, which ignores mode effects, and proposed methods, which adjust for mode effects under a simple measurement model, are investigated. The empirical/simulation work includes three studies. Two studies use a special type of data that include hypothetical true values. Since both studies include benchmark values, which may not be the usual case, a third study conducts an empirical comparison analysis for a case for which no benchmark values are available.