Please use this identifier to cite or link to this item:
|Title:||Contributions to estimation and interpretation of intervention effects and heterogeneity in meta-analysis|
Devereaux, Philip James
|Department:||Health Research Methodology|
|Keywords:||meta-analysis;heterogeneity;precision;estimation;Applied Statistics;Clinical Epidemiology;Statistical Methodology;Applied Statistics|
|Abstract:||<p><strong><em>Background and objectives</em></strong><strong> </strong></p> <p>Despite great statistical advances in meta-analysis methodology, most published meta-analyses make use of out-dated statistical methods and authors are unaware of the shortcomings associated with the widely employed methods. There is a need for statistical contributions to meta-analysis where: 1) improvements to current statistical practice in meta-analysis are conveyed at the level that most systematic review authors will be able to understand; and where: 2) current statistical methods that are widely applied in meta-analytic practice undergo thorough testing and examination. The objective of this thesis is to address some of this demand.</p> <p><strong><em>Methods</em></strong></p> <p>Four studies were conducted that would each meet one or both of the objectives. Simulation was used to explore the number of patients and events required to limit the risk of overestimation of intervention effects to ‘acceptable’ levels. Empirical assessment was used to explore the performance of the popular measure of heterogeneity, <em>I<sup>2</sup></em>, and its associated 95% confidence intervals (CIs) as evidence accumulates. Empirical assessment was also used to compare inferential agreement between the widely used DerSimonian-Laird random-effects model and four alternative models. Lastly, a narrative review was undertaken to identify and appraise available methods for combining health related quality of life (HRQL) outcomes.</p> <p><strong><em>Results and conclusion</em></strong></p> <p>The information required to limit the risk of overestimation of intervention effects is typically close to what is known as the optimal information size (OIS, i.e., the required meta-analysis sample size). <em>I<sup>2</sup> </em>estimates fluctuate considerably in meta-analyses with less than 15 trials and 500 events; their 95% confidence intervals provide desired coverage. The choice of random-effects has ignorable impact on the inferences about the intervention effect, but not on inferences about the degree of heterogeneity. Many approaches are available for pooling HRQL outcomes. Recommendations are provided to enhance interpretability. Overall, each manuscript met at least one thesis objective.</p>|
|Appears in Collections:||Open Access Dissertations and Theses|
Items in MacSphere are protected by copyright, with all rights reserved, unless otherwise indicated.