A METHODOLOGICAL STUDY EXAMINING APPROPRIATE CHOICE AND USE OF EFFECT MEASURES IN CLINICAL RESEARCH.

Abstract

This thesis addresses critical methodological challenges associated with the choice and interpretation of effect measures in clinical research, focusing on three interconnected themes that are related to effect sizes and include a) appropriate choice, interpretation, and implementation of binary and proportion effect measures, b) handling of heterogeneity of treatment effect in experimental studies and design-related bias in observational studies, and c) avoiding research waste through redefining conclusiveness in meta-analysis. First, the study re-evaluates the use and interpretation of commonly utilized binary effect measures, including the Odds Ratio (OR), Relative Risk (RR), and Risk Difference (RD). Noncollapsibility of the OR and collapsibility of the RR that have previously been discussed are reexamined to determine what this really means for clinical research. While noncollapsibility has been criticized, our research demonstrates that effect sizes that are noncollapsible have many distinct advantages over those that are collapsible, such as accurately reflecting outcome discrimination independently of baseline risks. Through theoretical analyses and empirical examples, we illustrate why the OR outperforms other binary measures within epidemiological analysis, especially when dealing with confounding and effect modification. This body of work also examined the proportion effect size with respect to different variance-stabilizing approaches for analysis commonly encountered in prevalence meta-analyses. Two main approaches were compared, which were the logit and Freeman-Tukey transformations, and extensive simulations were used alongside real-world meta-analytical data. Results clearly pinpoint the superiority of the Freeman-Tukey method with specific implementation that we describe over and above the logit transformation. The findings provide practical guidelines for researchers to select transformations tailored to their data characteristics, thus improving accuracy and reliability in meta-analytic prevalence estimation. Second, this body of work looks at experimental studies and the potential for heterogeneity of treatment effect related to prognostic covariates. The latter causes risk magnification in different subgroups, and the issue of covariate adjustment in randomized controlled trials (RCTs) was examined from this perspective. It was determined that covariate adjustment was useful for determining the closest empirical estimate of an individual treatment effect and also markedly addressed prognosis-based heterogeneity in meta-analysis that was independent of confounding. A second arm of this theme examined design-related bias in observational studies with a focus on immortal time bias in cohort studies. The latter is due to misclassification of follow-up time between groups, which is very commonly seen in cohort designs and has been addressed using time-redistribution methods. It was found that the common approach to time-redistribution, called the prescription time distribution method, did not sufficiently mitigate immortal time bias, and we propose, for the first time, an iterative time distribution method that resolves the problem seen earlier. Third, there has been an pressing need to redefine and create a standard for when an evidence synthesis through meta-analysis is conclusive such that further studies or meta-analysis on the topic are no longer required. Towards this aim, in this project, different indicators of conclusiveness in meta-analysis were reviewed and examined. None of these were found to be satisfactory, and a new methodology was defined and tested through simulation. This method resulted in the derivation of a meta-analytic convergence plot and an index of convergence, which was named the Doi-Abdulmajeed trial stability (DAts) index. Further, to improve applicability, the methodology has been programmed and implemented into the common statistical package Stata in the module exitma. Adoption of this methodology will now be the way forward to effectively minimize redundancy in research, thereby enhancing clinical decision-making and reducing research waste. In summary, this body of work aims to provide practical methodological recommendations designed to improve the validity and applicability of clinical research findings, ultimately enhancing evidence-based clinical practice.