Back to search

University of Ioannina

Exploring methodological challenges in network meta-analysis models and developing methodology for outlier detection

Abstract

dc:description

Systematic reviews and meta-analyses have been established as an integral part of comparativeeffectiveness research. The increasing number of different educational and psychological interventions in the educational system has led to the need for comparative effectiveness research with the aim to identify the best intervention. Network meta-analysis synthesizes both direct and indirect evidence, gives more powerful results and provides estimates with increased precision compared to pairwise estimates. NMA has become a popular statistical tool in evidence synthesis. Based on a database of published NMA from the onset until 14 April 2015, the time trend indicates the increasing number of published NMAs and the tendency for the use of appropriate methods. Moreover, the overview of the characteristics of published NMAs is a useful resource of information for methodologists that aim to update the current knowledge on appraising NMA methods. This collection of 456 published NMAs indicates that many NMAs provide important methodological limitations, but the comprehensive use of appropriate methodologies and completeness of reporting (such as the description of the statistical methods used) improved over the years. For example, anincreasing number of NMAs used appropriate methods to test the plausibility of the consistency assumption and in recent years around 90% of articles clearly reported whether a random-effects of the fixed-effect model was used. A common problem in the synthesis of studies is the existence of outlying or/and influential studies. Outlying and influential studies may bias the results but little work has been done for outlying identification in NMA. For this reason, this Thesis focuses on developing several methodologies for the identification of outliers and influential studies in network metaanalysis. Heterogeneity and inconsistency can be seen as differences in the potential effectmodifiers within and across the pairwise comparisons in a network of interventions. It is common that a potential source of heterogeneity and inconsistency is provided due to the existence of extreme study effects. Extreme study effects may be an outlier or influential study. A study that is far away from the rest of the data and does not explain by the assumed model defined as an outlier and a study that influences the model parameters (network estimates and heterogeneity estimator) defined as influential. Several methods for outlier and influential identification have been developed in a pairwise meta-analysis considering addition or deletion of studies, based on the likelihood or by taking alternative distributions for heterogeneity. In this dissertation, several simple measures for outlier and influential studies detection are provided. Measures considering the deletion of a study for outlier and influential studies detection are extended in NMA. A forward search algorithm, considered the addition of studies, has recently been developed in meta-regression. This algorithm starts with a subset of studies that considered outlying-free and it gradually adds studies until all studies entered. Sharp changes in monitoring measures during the search are considered potential outlying and/or influential studies. In this Thesis, the methodology with the forward search algorithm for outlying identification has been developed in the NMA model. Additionally, a novel model with shifting the variance taking into account outlying studies from meta-analysis to networkmeta-analysis model is extended. The advantage of the random shift variance model is that it offers the ability of down-weighting studies and therefore can be used as a sensitivity analysis. The several outlier and influential measures and two proposed methods in NMA for outlying identification, forward search algorithm and shift random NMA model, applied in real and in simulated datasets. Results of measures and methods indicate the potential source of outlying and influential cases in datasets. The methods are promising tools for the identification of outlying and influential cases and sources of heterogeneity and/or inconsistency. For the implementation of the several detection measures and methods a flexible and user-friendly software, an R package, called NMAoutlier, was developed with a description and details to provide guidance on how to use the R package through real datasets.

Degree

thesis:*
Grantor dc:publisher
University of Ioannina
Year dc:date
2020

Author and committee

dc:creator, dc:contributor.*
Authors dc:creator
  • Petropoulou, Maria
  • Πετροπούλου, Μαρία

Subjects

dc:subject × 18

Rights

Language dc:language
eng

Identifiers

dc:identifier.*
Identifier
10.12681/eadd/48001

Chain of custody

source
Harvested from
Greek National Archive of PhD Theses
Base URL
phdtheses.ekt.gr/eadd_oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
citation

Petropoulou, Maria; Πετροπούλου, Μαρία. Exploring methodological challenges in network meta-analysis models and developing methodology for outlier detection. University of Ioannina, 2020. http://hdl.handle.net/10442/hedi/48001