AMSTAR-2 (A Measurement Tool to Assess Systematic Reviews, version 2) and ROBIS are tools used to assess the methodological quality and the risk of bias in a systematic review (SR). We applied AMSTAR-2 and ROBIS to a sample of 200 published SRs. We investigated the overlap in their methodological constructs, responses by item, and overall, percentage agreement, direction of effect, and timing of assessments. AMSTAR-2 contains 16 items and ROBIS 24 items. Three items in AMSTAR-2 and nine in ROBIS did not overlap in construct. Of the 200 SRs, 73% were low or critically low quality using AMSTAR-2, and 81% had a high risk of bias using ROBIS. The median time to complete AMSTAR-2 and ROBIS was 51 and 64 minutes, respectively. When assessment times were calibrated to the number of items in each tool, each item took an average of 3.2 minutes per item for AMSTAR-2 compared to 2.7 minutes for ROBIS. Nine percent of SRs had opposing ratings (i.e., AMSTAR-2 was high quality while ROBIS was high risk). In both tools, three-quarters of items showed more than 70% agreement between raters after extensive training and piloting. AMSTAR-2 and ROBIS provide complementary rather than interchangeable assessments of systematic reviews. AMSTAR-2 may be preferable when efficiency is prioritized and methodological rigour is the focus, whereas ROBIS offers a deeper examination of potential biases and external validity. Given the widespread reliance on systematic reviews for policy and practice, selecting the appropriate appraisal tool remains crucial. Future research should explore strategies to integrate the strengths of both instruments while minimizing the burden on assessors.

The evaluation of the role of face masks in preventing respiratory infections is a paradigm case in synthesising complex evidence (i.e. extensive, diverse, technically specialised, and with multilevel chains of causality). Primary studies have assessed different mask types, diseases, populations, and settings using different research designs. Numerous review teams have attempted to synthesise this literature, in which observational (case–control, cohort, cross-sectional) and ecological studies predominate. Their findings and conclusions vary widely.

This article critically examines how 66 systematic reviews dealt with mask efficacy studies. Risk-of-bias tools produced unreliable assessments when—as was often the case—review teams lacked methodological expertise or topic-specific understanding. This was especially true when datasets were large and heterogeneous, with multiple biases playing out in different ways and requiring nuanced adjustments. In such circumstances, tools were sometimes used crudely and reductively rather than to support close reading of primary studies and guide expert judgments. Various moves by reviewers—excluding observational evidence altogether, assessing risk but not direction of biases, omitting distinguishing details of primary studies, and producing meta-analyses that combined studies of different designs or included studies at critical risk of bias—served to obscure important aspects of heterogeneity, resulting in bland and unhelpful summary statements.

We draw on philosophy to question the formulaic use of generic risk-of-bias tools, especially when the primary evidence demands expert understanding and tailoring of study quality questions to the topic. We call for more rigorous training and oversight of reviewers of complex evidence and for new review methods designed specifically for such evidence.

Critical appraisal is a core component of JBI qualitative evidence synthesis, offering insights into the quality of included studies and their potential influence on synthesized findings. However, limited guidance exists on whether, when, and how to exclude studies based on appraisal results. This study examined the methods used in JBI qualitative systematic reviews and the implications for synthesized findings. In this study, a systematic analysis of qualitative reviews published between 2018 and 2022 in JBI Evidence Synthesis was conducted. Data on decisions and their justifications were extracted from reviews and protocols. Descriptive and content analysis explored variations in the reported methods. Forty-five reviews were included. Approaches reported varied widely: 24% of reviews included all studies regardless of quality, while others applied exclusion criteria (36%), cutoff scores (11%), or multiple methods (9%). Limited justifications were provided for the approaches. Few reviews cited methodological references to support their decisions. Review authors reported their approach in various sections of the review, with inconsistencies identified in 18% of the sample. In addition, unclear or ambiguous descriptions were also identified in 18% of the included reviews. No clear differences were observed in ConQual scores between reviews that excluded studies and those that did not. Overall, the variability raises concerns about the credibility, transparency, and reproducibility of JBI qualitative systematic reviews. Decisions regarding the inclusion or exclusion of studies based on critical appraisal need to be clearly justified and consistently reported. Further methodological research is needed to support rigorous decision-making and to improve the reliability of synthesized findings.

Indirect treatment comparison (ITC) is widely used to estimate the comparative effectiveness of treatments when head-to-head trials are unavailable. For the typical scenario of anchored ITC where one trial compares drug A to drug C (AC trial) and another compares drug B to drug C (BC trial), the comparative effectiveness of drugs A versus B is calculated by subtracting (or dividing) the relative treatment effect of A versus C in the AC trial by that of B versus C in the BC trial, assuming the covariate distributions in both trials are balanced. This operation is valid only if the chosen effect measure is transitive, that is, in a three-arm randomized trial of drugs A, B, and C, the direct treatment effect of A versus B equals the indirect treatment effect of A versus B through their comparisons to C. For survival outcomes, many ITCs use the hazard ratio (HR) as the effect measure. In this article, we demonstrate that HR is generally not transitive and should be used with caution. As more reliable alternatives, we recommend effect measures with better transitivity properties: the restricted mean survival time (RMST) difference, the landmark survival probability difference (or ratio) at a prespecified time point, and the average hazard with survival weights (AH-SW) difference.

Bayesian hierarchical models offer a principled framework for adjusting for study-level bias in meta-analysis, but their complexity and sensitivity to prior specifications necessitate a systematic framework for robust application. This study demonstrates the application of a Bayesian workflow to this challenge, comparing a standard random-effects model to a bias-adjustment model across a real-world dataset and a targeted simulation study. The workflow revealed a high sensitivity of results to the prior on bias probability, showing that while the simpler random-effects model had superior predictive accuracy as measured by the widely applicable information criterion, the bias-adjustment model successfully propagated uncertainty by producing wider, more conservative credible intervals. The simulation confirmed the model’s ability to recover true parameters when priors were well-specified. These results establish the Bayesian workflow as a principled framework for diagnosing model sensitivities and ensuring the transparent application of complex bias-adjustment models in evidence synthesis.

A growing number of oncology treatments, such as bevacizumab, are used across multiple indications. However, in health technology assessment (HTA), their clinical and cost-effectiveness are typically appraised within a single target indication. This approach excludes a broader evidence base across other indications. To address this, we explored multi-indication meta-analysis methods that share evidence across indications.

We conducted a simulation study to evaluate alternative multi-indication synthesis models. This included univariate (mixture and non-mixture) methods synthesizing overall survival (OS) data and bivariate surrogacy models jointly modeling treatment effects on progression-free survival (PFS) and OS, pooling surrogacy parameters across indications. Simulated datasets were generated using a multistate disease progression model under various scenarios, including different levels of heterogeneity within and between indications, outlier indications, and varying data on OS for the target indication. We evaluated the performance of the synthesis models applied to the simulated datasets in terms of their ability to predict OS in a target indication.

The results showed univariate multi-indication methods could reduce uncertainty without increasing bias, particularly when OS data were available in the target indication. Compared with univariate methods, mixture models did not significantly improve performance and are not recommended for HTA. In scenarios where OS data in the target indication is absent and there are also outlier indications, bivariate surrogacy models showed promise in correcting bias relative to univariate models, though further research under realistic conditions is needed.

Multi-indication methods are more complex than traditional approaches but can potentially reduce uncertainty in HTA decisions.

Interest in large language models (LLMs) as a tool for meta-analyses and systematic reviews (MA/SRs) is growing. We prospectively developed 515 unique prompts by predefined screening-related categories and tested with open-access LLMs (Llama, Mistral) against four gold-standard MA/SRs from different medical fields published after the LLMs’ training cut-offs, using a Python-based pipeline. Heterogeneity between prompts was quantified, and hypothetical workload/cost reduction with top-performing prompts calculated. Across 12,360 pipeline runs, LLMs versus MA/SRs reached average recall/sensitivity = 83.6 ± 17.0%, precision = 18.5 ± 15.6%, specificity = 36.6 ± 23.7% F1-score = 27.6 ± 17.2%, and accuracy = 61.1 ± 11.0%. F1-scores were significantly higher when prompts focused on methods (0.78 ± 0.40%), explicitly mentioned MA/SR screening (0.81 ± 0.37%), included the comparison MA/SR’s title (5.64 ± 0.37%) or selection criteria (8.05 ± 0.68%), and with more LLM parameters (70b = 4.48 ± 0.31%, 123b = 7.77 ± 0.31%), but lower when screening abstracts instead of titles (−3.67 ± 0.28%). In LLM-base preselection, top-performing F1-score prompts (recall/sensitivity = 72.2%, specificity = 66.1%, precision = 28.6%) would reduce screening demands by 34.5%−37.5%, saving 8.4–8.8 weeks of work and 17,592–18,552. Recall/sensitivity increased with less MA/SR information contrasting F1-score results, which highlights a recall/sensitivity-precision/specificity trade-off. F1-score increased with detailed MA/SR information, while recall/sensitivity increased with shorter, zeroshot prompts. We provide the first prospectively assessed prompt engineering framework for early-stage LLM-based paper screening across medical fields. The publicly available Python pipeline and full prompt list used here support further development of LLM-based evidence synthesis.

1. Meta-analyses are a reliable method for a quantitative research synthesis. They are, however, prone to specific biases that can be introduced in the process. Such a bias could exist if primary literature produces similar results if coming from the same authors. Authorship network bias is the non-independence of effect sizes introduced by the overlap of authors of primary studies. If not accounted for, it can severely impact the quality of meta-analysis and the conclusions drawn from it.

2. To account for such non-independence, multilevel models with author clusters as an additional hierarchy level were recently suggested. We propose a new method for the detection of non-independent effect sizes based on authorship networks and for their correction.

3. An analysis of simulated data demonstrates the effectiveness of the here-suggested new method. We further applied our new method to nine exemplary meta-analyses.

4. Our new method for detection and effective correction can be easily integrated in existing meta-analysis workflows, using the functionality already offered by R’s metafor package.

5. Our goal is to enhance the reliability of meta-analyses by highlighting potential authorship network bias and offering a method to address this often-overlooked bias.

Meta-analytic confirmatory factor analysis (CFA) is a type of meta-analytic structural equation modeling (MASEM) that is useful for evaluating the factor structure of measurement scales based on data from multiple studies. Modeling the factor structure is just one example of the many potentially interesting research questions. Analyzing covariance matrices allows for the evaluation of measurement properties across studies, such as whether indicators are functioning the same across studies. For example, are some indicators more indicative of the common factor in certain types of studies than in others? The additional analysis of means of the observed variables opens up many other research questions to consider such as: “Are there mean differences in mental health between clinical and non-clinical samples?” To answer such questions, it is necessary to analyze both the covariance and the mean structure of the indicators. In this paper, we present, illustrate, and evaluate a method to incorporate the means of variables in the MASEM analyses of such datasets. We focus on meta-analytic CFA, with the aim of testing differences in latent means across studies. We provide illustrations of the comparison of latent means across groups of studies using two empirical datasets, for which data and analysis scripts are provided online. The performance of the new model was tested in a small-scale simulation study. The results showed adequate performance under the tested conditions. Finally, we discuss how the proposed method relates to other analysis options such as multigroup or multilevel structural equation modeling.

A key task in conducting systematic reviews is deduplicating the results from database searching. Deduplication using reference management software can be time-consuming and prone to error, while automated tools can be expensive and lack transparency. To support review teams, we evaluated eight deduplication tools: (1) The Automated Systematic Search Deduplicator (ASySD); (2) Covidence; (3) Deduklick; (4) EPPI-Reviewer; (5) PICO Portal; (6) Rayyan; (7) The Systematic Review Accelerator (SRA) Deduplicator: Focused; (8) The SRA Deduplicator: Relaxed. Five randomly selected Cochrane reviews had their searches rerun to create five gold standard sets. We compared the gold standard sets to the outputs of the eight deduplication tools and evaluated the results for: (1) unique records removed; (2) duplicate records retained; (3) time taken to deduplicate. Summed across all five reviews, the unique records removed in error ranged from 2 to 22. The three best tools were: (1) Rayyan; (2) Covidence; (3) SRA Deduplicator: Focused. The duplicate records retained in error ranged from 34 to 280, the three best tools were: (1) ASySD; (2) Rayyan; (3) EPPI-Reviewer. The time taken to deduplicate ranged from one minute to 20 hours and 34 minutes, the three fastest tools were: (1) SRA Deduplicator: Relaxed; (2) Deduklick; (3) Covidence. No tool performed so poorly that we don’t recommend using it. But, as all the tools had strengths and weaknesses, some are expensive while others require large amounts of manual checking time, we recommend review teams compare the tools across all three outcomes and choose the tool that best suits their needs.

Random-effects meta-analysis is a widely applied methodology to synthesize research findings of studies related to a specific scientific question. Besides estimating the mean effect, an important aim of the meta-analysis is to summarize the heterogeneity, that is, the variation in the underlying effects caused by the differences in study circumstances. The prediction interval is frequently used for this purpose: a 95% prediction interval contains the true effect of a similar new study in 95% of the cases when it is constructed, or in other words, it covers 95% of the true effects distribution on average in repeated sampling. In this article, after providing a clear mathematical background, we present an extensive simulation investigating the performance of all frequentist prediction interval methods published to date. The work focuses on the distribution of the coverage probabilities and how these distributions change depending on the amount of heterogeneity and the number of involved studies. Although the single requirement that a prediction interval has to fulfill is to keep a nominal coverage probability on average, we demonstrate why the distribution of coverages should not be disregarded. We show that for meta-analyses with small number of studies, this distribution has an unideal, asymmetric shape. We argue that assessing only the mean coverage can easily lead to misunderstanding and misinterpretation. The length of the intervals and the robustness of the methods concerning the non-normality of the true effects are also investigated.

The contrast-based model (CBM) is the most popular network meta-analysis (NMA) method, although alternative approaches, e.g., the baseline model (BM), have been proposed but seldom used. This article aims to illuminate the difference between the CBM and BM and explores when they produce different results. These models differ in key assumptions: The CBM assumes treatment contrasts are exchangeable across trials and models the reference (baseline) treatment’s outcome levels as fixed effects, while the BM further assumes that the baseline treatment’s outcome levels are exchangeable across trials and treats them as random effects. We show algebraically and graphically that the difference between the CBM and BM is analogous to the difference between the two analyses in a statistical conundrum called Lord’s Paradox, in which the t-test and analysis of covariance (ANCOVA) yield conflicting conclusions about the group difference in weight gain. We show that this conflict arises because the t-test compares the observed weight change, whereas ANCOVA compares an adjusted weight change. In NMA, analogously, the CBM compares observed treatment contrasts, while the BM compares adjusted treatment contrasts. We demonstrate how the difference in modeling baseline effects can cause the CBM and BM to give different results. The analogy of Lord’s Paradox provides insights into the different assumptions of the CBM and BM regarding the relationship between baseline effects and treatment contrasts. When these two models produce substantially different results, it may indicate a violation of the transitivity assumption. Therefore, we should be cautious in interpreting the results from either model.

A key output of network meta-analysis (NMA) is the relative ranking of treatments; nevertheless, it has attracted substantial criticism. Existing ranking methods often lack clear interpretability and fail to adequately account for uncertainty, overemphasizing small differences in treatment effects. We propose a novel framework to estimate treatment hierarchies in NMA using a probabilistic model, focusing on a clinically relevant treatment-choice criterion (TCC). Initially, we define a TCC based on smallest worthwhile differences (SWD), converting NMA relative treatment effects into treatment preference format. These data are then synthesized using a probabilistic ranking model, assigning each treatment a latent “ability” parameter, representing its propensity to yield clinically important and beneficial true treatment effects relative to the rest of the treatments in the network. Parameter estimation relies on the maximum likelihood theory, with standard errors derived asymptotically from the Hessian matrix. To facilitate the use of our methods, we launched the R package mtrank. We applied our method to two clinical datasets: one comparing 18 antidepressants for major depression and another comparing 6 antihypertensives for the incidence of diabetes. Our approach provided robust, interpretable treatment hierarchies that account for a concrete TCC. We further examined the agreement between the proposed method and existing ranking metrics in 153 published networks, concluding that the degree of agreement depends on the precision of the NMA estimates. Our framework offers a valuable alternative for NMA treatment ranking, mitigating overinterpretation of minor differences. This enables more reliable and clinically meaningful treatment hierarchies.

Individual patient data meta-analyses (IPDMAs) provide powerful tools for synthesizing evidence across studies, yet methods for addressing unmeasured confounding in observational IPDMAs with survival outcomes are rarely implemented. Instrumental variable (IV) approaches offer causal inference capabilities but face practical challenges in hierarchical data structures, particularly the lack of standard diagnostics for instrument strength in nonlinear mixed-effects models. We adapt and evaluate a frequentist mixed-effects two-stage residual inclusion (2SRI) framework for survival IPDMAs, extending traditional IV methods to accommodate study-level and temporal clustering while handling time-to-event outcomes through Cox proportional hazards models. Because classical F-statistics are unavailable for logistic mixed-effects first-stage models, we propose the Wald $\chi ^2$ statistic as a practical instrument-strength diagnostic and empirically characterize its relationship to estimator performance. Through a comprehensive simulation study with 48 scenarios—varying unmeasured confounding (weak to very strong), instrument–treatment association strength (0.3–1.0), and cross-study IV allocation patterns—we evaluated 2SRI against naive mixed-effects Cox models using bias, coverage, variance, and mean squared error. The design was anchored to realistic IPDMA structure (10 studies, $N \approx 4,357$) from pooled Ebola data, with 1,000 replications per scenario. Results show that under weak confounding, naive models dominate on all metrics. With moderate-to-strong confounding and realized Wald $\chi ^2$ exceeding 150–200, mixed-effects 2SRI substantially reduces bias and achieves near-nominal coverage, though with inflated variance. We provide empirical guideposts linking realized first-stage strength to expected performance, enabling analysts to judge when 2SRI will outperform conventional approaches in hierarchical survival IPDMAs. All simulations assume a common treatment effect across studies. Performance under heterogeneous effects remains to be established.

Ranking multiple interventions is a crucial task in network meta-analysis (NMA) to guide clinical and policy decisions. However, conventional ranking methods often oversimplify treatment distinctions, potentially yielding misleading conclusions due to inherent uncertainty in relative intervention effects. To address these limitations, we propose a novel Bayesian rank-clustering estimation approach, termed rank-clustering estimation (RaCE), specifically developed for NMA. Rather than identifying a single “best” intervention, RaCE enables the probabilistic clustering of interventions with similar effectiveness, offering a more nuanced and parsimonious interpretation. By decoupling the clustering procedure from the NMA modeling process, RaCE is a flexible and broadly applicable approach that can accommodate different types of outcomes (binary, continuous, and survival), modeling approaches (arm-based and contrast-based), and estimation frameworks (frequentist or Bayesian). Simulation studies demonstrate that RaCE effectively captures rank-clusters even under conditions of substantial uncertainty and overlapping intervention effects, providing more reasonable result interpretation than traditional single-ranking methods. We illustrate the practical utility of RaCE through an NMA application to frontline immunochemotherapies for follicular lymphoma, revealing clinically relevant clusters among treatments previously assumed to have distinct ranks. Overall, RaCE provides a valuable tool for researchers to enhance rank estimation and interpretability, facilitating evidence-based decision-making in complex intervention landscapes.

This study analyzes 1,000 meta-analyses drawn from 10 disciplines—including medicine, psychology, education, biology, and economics—to document and compare methodological practices across fields. We find large differences in the size of meta-analyses, the number of effect sizes per study, and the types of effect sizes used. Disciplines also vary in their use of unpublished studies, the frequency and type of tests for publication bias, and whether they attempt to correct for it. Notably, many meta-analyses include multiple effect sizes from the same study, yet fail to account for statistical dependence in their analyses. We document the limited use of advanced methods—such as multilevel models and cluster-adjusted standard errors—that can accommodate dependent data structures. Correlations are frequently used as effect sizes in some disciplines, yet researchers often fail to address the methodological issues this introduces, including biased weighting and misleading tests for publication bias. We also find that meta-regression is underutilized, even when sample sizes are large enough to support it. This work serves as a resource for researchers conducting their first meta-analyses, as a benchmark for researchers designing simulation experiments, and as a reference for applied meta-analysts aiming to improve their methodological practices.

Systematic reviews are often characterized as being inherently replicable, but several studies have challenged this claim. The objective of the study was to investigate the variation in results following independent replication of literature searches and meta-analyses of systematic reviews. We included 10 systematic reviews of the effects of health interventions published in November 2020. Two information specialists repeated the original database search strategies. Two experienced review authors screened full-text articles, extracted data, and calculated the results for the first reported meta-analysis. All replicators were initially blinded to the results of the original review. A meta-analysis was considered not ‘fully replicable’ if the original and replicated summary estimate or confidence interval width differed by more than 10%, and meaningfully different if there was a difference in the direction or statistical significance. The difference between the number of records retrieved by the original reviewers and the information specialists exceeded 10% in 25/43 (58%) searches for the first replicator and 21/43 (49%) searches for the second. Eight meta-analyses (80%, 95% CI: 49–96) were initially classified as not fully replicable. After screening and data discrepancies were addressed, the number of meta-analyses classified as not fully replicable decreased to five (50%, 95% CI: 24–76). Differences were classified as meaningful in one blinded replication (10%, 95% CI: 1–40) and none of the unblinded replications (0%, 95% CI: 0–28). The results of systematic review processes were not always consistent when their reported methods were repeated. However, these inconsistencies seldom affected summary estimates from meta-analyses in a meaningful way.

Externally controlled single-arm trials are critical to assess treatment efficacy across therapeutic indications for which randomized controlled trials are not feasible. A closely-related research design, the unanchored indirect treatment comparison, is often required for disconnected treatment networks in health technology assessment. We present a unified causal inference framework for both research designs. We develop an estimator that augments a popular weighting approach based on entropy balancing—matching-adjusted indirect comparison (MAIC)—by fitting a model for the conditional outcome expectation. The predictions of the outcome model are combined with the entropy balancing MAIC weights. While the standard MAIC estimator is singly robust where the outcome model is non-linear, our augmented MAIC approach is doubly robust (DR), providing increased robustness against model misspecification. This is demonstrated in a simulation study with binary outcomes and a logistic outcome model, where the augmented estimator demonstrates its DR property, while exhibiting higher precision than all non-augmented weighting estimators and near-identical precision to G-computation. We describe the extension of our estimator to the setting with unavailable individual participant data for the external control, illustrating it through an applied example. Our findings reinforce the understanding that entropy balancing-based approaches have desirable properties compared to standard “modeling” approaches to weighting, but should be augmented to improve protection against bias and guarantee double robustness.

Meta-analysis synthesizes evidence from multiple randomized clinical trials and informs evidence-based practices across various medical domains. Recently, causally interpretable meta-analysis has been proposed and applied to treatment evaluations for target populations, requiring individual participant data (IPD). Standard meta-analysis assumes transportability or exchangeability of a (conditional) relative effect (such as relative risk or odds ratio), which may be violated when the relative effects are correlated with the baseline risks across clinical trials. In addition, the weighted average of some study-specific effect measures such as the (log) odds ratios or the (log) hazard ratios is non-collapsible and does not correspond to any target population. Furthermore, when the randomization ratios between treated versus untreated arms vary across trials, confounding bias may occur. To address these challenges, we propose a causal meta-analysis (CMA) framework using only aggregated data, enabling causally interpretable and accurate estimation for different target populations. The CMA adjusts its weights for treatment effect across various target populations, including the average treatment effect (ATE), the ATE on the treated (ATT) population, the ATE on the control (ATC) population, and the ATE in the overlap (ATO) population. Mathematically, we discover the connection between traditional meta-analysis estimators and CMAs. For example, the Mantel–Haenszel weighted meta-analysis is equivalent to the CMA with ATO.