ISR Awards

Doctoral Dissertation Research: Investigating the Bias of Alternative Statistical Inference Methods in Sequential Mixed-Mode Surveys

In a sequential mixed-mode survey design, a combination of data collection modes is used sequentially to reduce nonresponse bias under certain cost constraints. However, nonrandom mixes of modes yield unknown bias properties for the population estimates such as means and totals. This project aims to develop statistical inference methods accounting for both nonresponse and nonrandom response effects. Additionally, response effects by mode will be evaluated by analytically controlling the demographics and socioeconomic characteristics of the persons who respond via different modes. The American Community Survey (ACS), conducted by the Census Bureau, is the most prominent survey that uses such a sequential mixed-mode survey design, and therefore is interesting as a test bed for the empirical evaluations. While the analytical methods discussed in this proposal are also applicable to the other items, the analyses will concentrate on two variables of interest: (1) personal income, and (2) health insurance coverage. Both income and health insurance coverage data are important in public policy and decision making, and the government survey data are the only available data in some instances. The mode effects literature suggests that both of these variables are subject to mode effects. Thus, the proposed methods are expected to improve the comparability of the estimates from the different modes while incorporating nonresponse adjustment and controlling for nonrandom response effects.
The intellectual merit of the research is the development of an inferential method that adjusts for both nonresponse and nonrandom response effects in the context of sequential mixed-mode surveys. In the presence of nonignorable response effects by mode, the bias properties for the population estimates are not known and the existing inferential methods do not control for the nonrandom response effects by mode. The broader impact from the study is that it will inform survey researchers as to whether adjustments are required for sequential mixed-mode survey data inference, from a total survey error perspective. Findings of significant differences between response modes and smaller relative biases yielded by the developed method would inform researchers that alternative statistical inference methods need be considered. Finding of minimal differences, by contrast, would suggest the control of nonrandom response effects is not required and the varying response effects by mode could be ignorable in the context of sequential mixed-mode survey inference.