Graduation Year

2003

Document Type

Dissertation

Degree

Ph.D.

Degree Granting Department

Psychology

Major Professor

Michael Brannick, Ph. D.

Committee Member

Walter Borman, Ph.D

Committee Member

Carnot Nelson, Ph.D

Committee Member

Billy N. Kinder, Ph.D

Committee Member

Joel Thompson Ph.D

Keywords

statistics, testing methodology, validity, r to z transformation, moderators

Abstract

In the last few years, several studies have attempted to meta-analyze reliability estimates. The initial study, to outline a methodology for meta-analyzing reliability coefficients, was published by Vacha-Haase in 1998. Vacha-Haase used a very basic meta-analytic model to find a mean effect size (reliability) across studies. There are two main reasons for meta-analyzing reliability coefficients. First, recent research has shown that many studies fail to report the appropriate reliability for the measure and population of the actual study (Vacha-Haase, Ness, Nilsson and Reetz, 1999; Whittington, 1998; Yin and Fan, 2000). Second, very little research has been published describing the way reliabilities for the same measure vary according to moderators such as time, form length, population differences in trait variability and others. Vacha-Haase (1998) proposed meta-analysis, as a method by which the impact of moderators may become better understood.

Although other researchers have followed the Vacha-Haase example and meta-analyzed the reliabilities for several measures, little has been written about the best methodology to use for such analysis. Reliabilities are much larger on average than are validities, and thus tend to show greater skew in their sampling distributions. This study took a closer look at the methodology with which reliability can be meta-analyzed. Specifically, a Monte Carlo study was run so that population characteristics were known. This provided a unique ability to test how well each of three methods estimates the true population characteristics. The three methods studied were the Vacha-Haase method as outlined in her 1998 article, the well-known Hunter and Schmidt "bare bones method" (1990) and the random-effects version of Hedges' method as described by Lipsey and Wilson (2001). The methods differ both in how they estimate the random-effects variance component (or in one case, whether the random-effects variance component is estimated at all) and in how they treat moderator variables. Results showed which of these methods is best applied to reliability meta-analysis. A combination of the Hunter and Schmidt (1999) method and weighted least squares regression is proposed.

Download

Share

COinS