Research Synthesis Methods最新文献

Collecting data for an individual participant data meta-analysis (IPDMA) project can be time consuming and resource intensive and could still have insufficient power to answer the question of interest. Therefore, researchers should consider the power of their planned IPDMA before collecting IPD. Here we propose a method to estimate the power of a planned IPDMA project aiming to synthesise multiple cohort studies to investigate the (unadjusted or adjusted) effects of potential prognostic factors for a binary outcome. We consider both binary and continuous factors and provide a three-step approach to estimating the power in advance of collecting IPD, under an assumption of the true prognostic effect of each factor of interest. The first step uses routinely available (published) aggregate data for each study to approximate Fisher's information matrix and thereby estimate the anticipated variance of the unadjusted prognostic factor effect in each study. These variances are then used in step 2 to estimate the anticipated variance of the summary prognostic effect from the IPDMA. Finally, step 3 uses this variance to estimate the corresponding IPDMA power, based on a two-sided Wald test and the assumed true effect. Extensions are provided to adjust the power calculation for the presence of additional covariates correlated with the prognostic factor of interest (by using a variance inflation factor) and to allow for between-study heterogeneity in prognostic effects. An example is provided for illustration, and Stata code is supplied to enable researchers to implement the method.

The application of network meta-analysis is becoming increasingly widespread, and for a successful implementation, it requires that the direct comparison result and the indirect comparison result should be consistent. Because of this, a proper detection of inconsistency is often a key issue in network meta-analysis as whether the results can be reliably used as a clinical guidance. Among the existing methods for detecting inconsistency, two commonly used models are the design-by-treatment interaction model and the side-splitting models. While the original side-splitting model was initially estimated using a Bayesian approach, in this context, we employ the frequentist approach. In this paper, we review these two types of models comprehensively as well as explore their relationship by treating the data structure of network meta-analysis as missing data and parameterizing the potential complete data for each model. Through both analytical and numerical studies, we verify that the side-splitting models are specific instances of the design-by-treatment interaction model, incorporating additional assumptions or under certain data structure. Moreover, the design-by-treatment interaction model exhibits robust performance across different data structures on inconsistency detection compared to the side-splitting models. Finally, as a practical guidance for inconsistency detection, we recommend utilizing the design-by-treatment interaction model when there is a lack of information about the potential location of inconsistency. By contrast, the side-splitting models can serve as a supplementary method especially when the number of studies in each design is small, enabling a comprehensive assessment of inconsistency from both global and local perspectives.

Multi-lab projects are large scale collaborations between participating data collection sites that gather empirical evidence and (usually) analyze that evidence using meta-analyses. They are a valuable form of scientific collaboration, produce outstanding data sets and are a great resource for third-party researchers. Their data may be reanalyzed and used in research synthesis. Their repositories and code could provide guidance to future projects of this kind. But, while multi-labs are similar in their structure and aggregate their data using meta-analyses, they deploy a variety of different solutions regarding the storage structure in the repositories, the way the (analysis) code is structured and the file-formats they provide. Continuing this trend implies that anyone who wants to work with data from multiple of these projects, or combine their datasets, is faced with an ever-increasing complexity. Some of that complexity could be avoided. Here, we introduce MetaPipeX, a standardized framework to harmonize, document and analyze multi-lab data. It features a pipeline conceptualization of the analysis and documentation process, an R-package that implements both and a Shiny App (https://www.apps.meta-rep.lmu.de/metapipex/) that allows users to explore and visualize these data sets. We introduce the framework by describing its components and applying it to a practical example. Engaging with this form of collaboration and integrating it further into research practice will certainly be beneficial to quantitative sciences and we hope the framework provides a structure and tools to reduce effort for anyone who creates, re-uses, harmonizes or learns about multi-lab replication projects.

Accurate data extraction is a key component of evidence synthesis and critical to valid results. The advent of publicly available large language models (LLMs) has generated interest in these tools for evidence synthesis and created uncertainty about the choice of LLM. We compare the performance of two widely available LLMs (Claude 2 and GPT-4) for extracting pre-specified data elements from 10 published articles included in a previously completed systematic review. We use prompts and full study PDFs to compare the outputs from the browser versions of Claude 2 and GPT-4. GPT-4 required use of a third-party plugin to upload and parse PDFs. Accuracy was high for Claude 2 (96.3%). The accuracy of GPT-4 with the plug-in was lower (68.8%); however, most of the errors were due to the plug-in. Both LLMs correctly recognized when prespecified data elements were missing from the source PDF and generated correct information for data elements that were not reported explicitly in the articles. A secondary analysis demonstrated that, when provided selected text from the PDFs, Claude 2 and GPT-4 accurately extracted 98.7% and 100% of the data elements, respectively. Limitations include the narrow scope of the study PDFs used, that prompt development was completed using only Claude 2, and that we cannot guarantee the open-source articles were not used to train the LLMs. This study highlights the potential for LLMs to revolutionize data extraction but underscores the importance of accurate PDF parsing. For now, it remains essential for a human investigator to validate LLM extractions.

Citation indices providing information on backward citation (BWC) and forward citation (FWC) links are essential for literature discovery, bibliographic analysis, and knowledge synthesis, especially when language barriers impede document identification. However, the suitability of citation indices varies. While some have been analyzed, the majority, whether new or established, lack comprehensive evaluation. Therefore, this study evaluates the citation coverage of the citation indices of 59 databases, encompassing the widely used Google Scholar, Scopus, and Web of Science alongside many others never previously analyzed, such as the emerging Lens, Scite, Dimensions, and OpenAlex or the subject-specific PubMed and JSTOR. Through a comprehensive analysis using 259 journal articles from across disciplines, this research aims to guide scholars in selecting indices with broader document coverage and more accurate and comprehensive backward and forward citation links. Key findings highlight Google Scholar, ResearchGate, Semantic Scholar, and Lens as leading options for FWC searching, with Lens providing superior download capabilities. For BWC searching, the Web of Science Core Collection can be recommended over Scopus for accuracy. BWC information from publisher databases such as IEEE Xplore or ScienceDirect was generally found to be the most accurate, yet only available for a limited number of articles. The findings will help scholars conducting systematic reviews, meta-analyses, and bibliometric analyses to select the most suitable databases for citation searching.

Meta-analyses examining dichotomous outcomes often include single-zero studies, where no events occur in intervention or control groups. These pose challenges, and several methods have been proposed to address them. A fixed continuity correction method has been shown to bias estimates, but it is frequently used because sometimes software (e.g., RevMan software in Cochrane reviews) uses it as a default. We aimed to empirically compare results using the continuity correction with those using alternative models that do not require correction. To this aim, we reanalyzed the original data from 885 meta-analyses in Cochrane reviews using the following methods: (i) Mantel–Haenszel model with a fixed continuity correction, (ii) random effects inverse variance model with a fixed continuity correction, (iii) Peto method (the three models available in RevMan), (iv) random effects inverse variance model with the treatment arm continuity correction, (v) Mantel–Haenszel model without correction, (vi) logistic regression, and (vii) a Bayesian random effects model with binominal likelihood. For each meta-analysis we calculated ratios of odds ratios between all methods, to assess how the choice of method may impact results. Ratios of odds ratios <0.8 or <1.25 were seen in ~30% of the existing meta-analyses when comparing results between Mantel–Haenszel model with a fixed continuity correction and either Mantel–Haenszel model without correction or logistic regression. We concluded that injudicious use of the fixed continuity correction in existing Cochrane reviews may have substantially influenced effect estimates in some cases. Future updates of RevMan should incorporate less biased statistical methods.

When studies use different scales to measure continuous outcomes, standardised mean differences (SMD) are required to meta-analyse the data. However, outcomes are often reported as endpoint or change from baseline scores. Combining corresponding SMDs can be problematic and available guidance advises against this practice. We aimed to examine the impact of combining the two types of SMD in meta-analyses of depression severity. We used individual participant data on pharmacological interventions (89 studies, 27,409 participants) and internet-delivered cognitive behavioural therapy (iCBT; 61 studies, 13,687 participants) for depression to compare endpoint and change from baseline SMDs at the study level. Next, we performed pairwise (PWMA) and network meta-analyses (NMA) using endpoint SMDs, change from baseline SMDs, or a mixture of the two. Study-specific SMDs calculated from endpoint and change from baseline data were largely similar, although for iCBT interventions 25% of the studies at 3 months were associated with important differences between study-specific SMDs (median 0.01, IQR −0.10, 0.13) especially in smaller trials with baseline imbalances. However, when pooled, the differences between endpoint and change SMDs were negligible. Pooling only the more favourable of the two SMDs did not materially affect meta-analyses, resulting in differences of pooled SMDs up to 0.05 and 0.13 in the pharmacological and iCBT datasets, respectively. Our findings have implications for meta-analyses in depression, where we showed that the choice between endpoint and change scores for estimating SMDs had immaterial impact on summary meta-analytic estimates. Future studies should replicate and extend our analyses to fields other than depression.