Abstract/Summary

There is growing awareness of the importance of effect sizes in clinical trials. Effect sizes determine the practical significance of results and play an important role in power analysis and meta-analysis. While effect size estimation methods have been extensively studied within the univariate domain, comparatively estimating effect sizes in trials with multivariate responses such as repeated measures has received little attention in the literature. In addition, univariate effect sizes are inadequate for quantifying overall group differences in multivariate data. Although there are a few effect size indices that are appropriate for a repeated measures design, the reporting of effect sizes presents a challenge since some are either difficult to calculate or interpret, or are relatively uncommon. In this thesis, novel measures of overall effect size in a repeated measures design, which are consistent and efficient and explicitly account for the correlation structure that is inherent in multivariate responses, are developed and evaluated. Additionally, the accuracy, precision and efficiency of the proposed effect sizes are compared with common effect sizes for repeated measures from two different families, standardised mean difference and proportion of variance explained, via extensive simulation studies. Results from across a wide range of conditions revealed that the proposed effect sizes compare favourably. In particular, they outperformed all other effect sizes in trials involving small sample sizes or small population effect sizes while the common existing effect sizes can produce inaccurate or inefficient inference. The proposed effect sizes are easy to calculate and interpret, and are comparable across studies to facilitate meta-analysis. Furthermore, extensive simulation evaluation show that the proposed effect sizes are robust against variance heterogeneity. The proposed effect sizes provide a more trust worthy measure of overall effect size and are recommended for further use.