In July 2018, the RSS South West local group enjoyed three talks on statistical methods for evidence synthesis in health. The central theme of these talks was about methodological advances in meta-analysis and their applications to inform personalised treatment decisions. Two talks took place at the University of Plymouth and one at the University of Exeter.
X-META: a comprehensive toolbox for advanced meta-analysis
Dr Yong Chen of the University of Pennsylvania discussed a wide range of advanced meta-analysis methods and their implementation in X-META, a toolbox developed by his team at the Perelman School of Medicine. Yong started with a brief history of meta-analysis, from more than one hundred years ago when Karl Pearson established the first formal data pooling technique, to about forty years ago when the term meta-analysis was coined by the psychologist Gene Glass. Yong then reviewed present day multivariate meta-analysis methods, including pseudo and restricted maximum likelihood techniques, and multivariate methods of moment. He then discussed new marginal methods of moment technique and Riley’s model, both of which do not require information about the often missing within-study correlation. He compared these methods using several examples, illustrating the benefit of borrowing information when the correlation between outcomes is strong. Next, Yong discussed hybrid Bayesian hierarchical and composite likelihood methods for diagnostic test meta-analyses. Publication bias is a common issue in meta-analysis, with funnel plots and Egger’s regression test being used to assess the association between effect measures and standard errors. However, these techniques have not been developed for the situations when there are multiple outcomes. Yong discussed 'Galaxy plot', a new visualisation tool to estimate the pooled effect size while correcting for publication bias in multivariate meta-analysis.
In the second part of his talk, Yong discussed network meta-analysis, a technique to combine evidence from direct and indirect comparisons that allows multiple treatments to be compared and ranked. Yong discussed an example network meta-analysis, which included 68 randomised controlled trials comparing 14 antimanic drugs for acute mania. He illustrated how the ranking of antimanic drugs was visualised in the current literature by considering both the efficacy and the safety of the drugs. He then discussed how patients’ preference about efficacy and safety can be included in the analyses to inform personalised medical decision-making.
The third part of Yong’s talk was about work on automating systematic review process. A large number of published studies often result from literature searches on databases. Subsequently, many abstracts and complete articles have to be studied to decide on their relevance to research questions. Yong discussed methods to move from a manual review to a semi-automated procedure, which on one hand speeds up the systematic review process, and on the other hand, takes account of content experts’ judgement on the relevance of the studies.
Recent advances in network meta-analysis
Dr Orestis Efthimiou of the University of Bern discussed three main research projects which he is currently working on in the area of network meta-analysis.
There is growing interest in including both randomised controlled trials and non-randomised studies in network meta-analysis to provide more precise inference. The first project Orestis discussed is about three approaches for including non-randomised studies in network meta-analysis. The first was design-adjusted analysis, where study-specific bias and variance-inflated factors were incorporated in the network meta-analysis model to take account of the risk of bias in each study’s results. In the second approach, data from non-randomised studies were first meta-analysed separately. Subsequently, these results informed prior distributions for a Bayesian network meta-analysis. In the third approach, hierarchical models were used to account for different study designs. Orestis showed that in a network meta-analysis comparing fifteen pharmacological treatments and a placebo for schizophrenia, the inclusion of non-randomised studies helped improve the connectivity of the network of treatments. In design-adjusted analysis, the contributions from non-randomised studies were down weighted, compared to naïve analyses where the potential bias from non-randomised studies is ignored.
The second project Orestis discussed was about individual participant data (IPD) network meta-analysis to inform personalised treatment decision. Orestis presented a case study comparing three treatments, psychotherapy, drugs, and a combination of psychotherapy and drugs, for treating chronic depression. A repeated-measurements IPD network meta-regression was used to account for prognostic factors and effect modifiers, as well as to take into account correlations between multiple measurements within participants. Orestis presented estimated treatment effects for subgroups of patients with similar characteristics. He then commented that numerous tables are often needed when there are multiple outcomes of interest, making the results hard to read. In order to overcome this, Orestis introduced a web application to help making prediction about the best treatments for individual participants.
The third part of Orestis’ talk was about network meta-analysis of binary outcomes in the situation when the event of interest is rare. The commonly used odds ratio and its standard error may not be defined when some events do not occur. Orestis presented a generalisation of the Mantel-Haenszel methods for estimating odds ratios for rare events in network meta-analysis. The extended method was illustrated by an application to compare treatments to decrease blood loss in liver transplantation. Orestis cautioned that the concept of statistical significance becomes problematic, and that when performing Bayesian analyses, the choice between different "non-informative" prior distributions may have an influence on results. Sensitivity analyses are thus needed to assess the robustness of conclusions.
Network meta-analysis of N-of-1 trials
Professor Christopher Schmid of Brown University discussed statistical methods for combining results from multiple clinical trials, with each consisting of one participant. Christopher started by discussing the gap between research knowledge gained from clinical trials and their direct relevance to local clinical practice. In particular, the average effects estimated from randomised controlled trials may not directly apply to all individuals. The net treatment benefits could be positive for some individuals, but zero or even negative for others. A N-of-1 trial is a case study of a single participant. Using crossover designs, the participant serves as their own control to compare the efficacy of interventions, typically for chronic diseases.
Christopher discussed the social benefits of N-of-1 trials, including their ability to engage patients and clinicians in shared-decision making and to enhance scientific literacy in the general public. Christopher then discussed examples of funded N-of-1 studies, from a single N-of-1 trial about Fibromyalgia to multiple N-of-1 trials comparing two (the PRODUCE study) or multiple (the PREEMPT study) treatments. He then discussed using graphs and probability statements to present analysis results to stakeholders, and the need to customize the presentations according to their feedback.
In N-of-1 trials, repeated measurements on the participant are taken over time due to the cross-over design. There are potential correlations between the measurements within each participant. Christopher discussed the development of statistical models, from simple linear models for estimating treatment and time trend effect, to more complex models incorporating correlation in repeated measures using autoregressive structures. The methodology was extended to hierarchical models of multiple N-of-1 trials and finally to network meta-analysis of multiple N-of-1 trials. The results from network meta-analysis show an increased precision in the estimated treatment effects compared to individual trial analysis. At last, probabilities of treatment effectiveness were graphically presented for each individual patient, providing informative recommendation for personalising treatment choice.