Evaluation of Modern Missing Data Handling Methods for Coefficient Alpha

Authors

Katerina Matysova, University of Nebraska - LincolnFollow

First Advisor

Rafael De Ayala

Date of this Version

Fall 12-2019

Comments

A THESIS Presented to the Faculty of the Graduate College at the University of Nebraska in Partial Fulfillments of Requirements for the Degree of Master of Arts, Major: Educational Psychology, Under the Supervision of Professor Rafael De Ayala. Lincoln, Nebraska: December, 2019

Copyright 2019 Katerina Matysova

Abstract

When assessing a certain characteristic or trait using a multiple item measure, quality of that measure can be assessed by examining the reliability. To avoid multiple time points, reliability can be represented by internal consistency, which is most commonly calculated using Cronbach’s coefficient alpha. Almost every time human participants are involved in research, there is missing data involved. Missing data means that even though complete data were expected to be collected, some data are missing. Missing data can follow different patterns as well as be the result of different mechanisms. One traditional way to deal with missing data is listwise deletion, in which every observation with at least one missing value is discarded. Modern missing data techniques include multiple imputation and maximum likelihood estimation, which use the observed data to create an estimate for the missing values in order to utilize the whole sample size. The present study sought to examine the effect of missing data on coefficient alpha under certain conditions as well as to compare multiple imputation to listwise deletion in its effectiveness to handle missing data across those conditions. The results indicated that coefficient alpha is sensitive to numerous factors in the presence of missing data such as reliability level, sample size, missing data percentage, and missing data mechanism. As expected, there was little difference between listwise deletion and multiple imputation when data were missing completely at random, but multiple imputation performed better when data were missing at random and missing not at random. While listwise deletion always underestimated the true reliability, multiple imputation only underestimated the true reliability when data were missing not at random.

Advisor: Rafael De Ayala