Date of Completion

8-22-2018

Embargo Period

8-22-2018

Keywords

Missing Data; Multiple Imputation; Power Calculation; Rates of missing information; Sample Size

Major Advisor

Ofer Harel

Associate Advisor

Haim Bar

Associate Advisor

Elizabeth Schifano

Field of Study

Statistics

Degree

Doctor of Philosophy

Open Access

Open Access

Abstract

While Multiple Imputation (MI) has become one of the most broadly used methods for handling incomplete data, many questions remain unanswered regarding statistical inference when MI is used with incomplete data. One such question is how to calculate statistical power. Although it is widely acknowledged that MI improves estimation efficiency, reduces estimation bias, and partially restores power loss, there is a gap in the literature as far as quantifying the power gained from using MI over complete case analysis (CCA). Furthermore, the rates of missing information are well developed for traditional MI, but not for newly-adjusted MI. This thesis presents methodologies and simulation studies to calculate statistical power when MI is used, to compare the performance of MI with that of CCA, and to examine under which conditions MI can better restore statistical power. We also provide formulas to compute the rates of missing information for an adjusted two-stage MI, and apply them to evaluate the impact of an extra information source.

Download
COinS