Researchers would often like to leverage data from a collection of sources (e.g., meta-analyses of randomized trials, multi-center trials, pooled analyses of observational cohorts) to estimate causal effects in a target population of interest. However, because different data sources typically represent different underlying populations, traditional meta-analytic methods may not produce causally interpretable estimates that apply to any reasonable target population. In this article, we present the CausalMetaR R package, which implements robust and efficient methods to estimate causal effects in a given internal or external target population using multi-source data. The package includes estimators of average and subgroup treatment effects for the entire target population. To produce efficient and robust estimates of causal effects, the package implements doubly robust and non-parametric efficient estimators and supports using flexible data-adaptive (e.g., machine learning techniques) methods and cross-fitting techniques to estimate the nuisance models (e.g., the treatment model, the outcome model). We briefly review the methods, describe the key features of the package, and demonstrate how to use the package through an example. The package aims to facilitate causal analyses in the context of meta-analysis.

Systematic reviews are essential for evidence-based health care, but conducting them is time- and resource-consuming. To date, efforts have been made to accelerate and (semi-)automate various steps of systematic reviews through the use of artificial intelligence (AI) and the emergence of large language models (LLMs) promises further opportunities. One crucial but complex task within systematic review conduct is assessing the risk of bias (RoB) of included studies. Therefore, the aim of this study was to test the LLM Claude 2 for RoB assessment of 100 randomized controlled trials, published in English language from 2013 onwards, using the revised Cochrane risk of bias tool (‘RoB 2’; involving judgements for five specific domains and an overall judgement). We assessed the agreement of RoB judgements by Claude with human judgements published in Cochrane reviews. The observed agreement between Claude and Cochrane authors ranged from 41% for the overall judgement to 71% for domain 4 (‘outcome measurement’). Cohen’s κ was lowest for domain 5 (‘selective reporting’; 0.10 (95% confidence interval (CI): −0.10–0.31)) and highest for domain 3 (‘missing data’; 0.31 (95% CI: 0.10–0.52)), indicating slight to fair agreement. Fair agreement was found for the overall judgement (Cohen’s κ: 0.22 (95% CI: 0.06–0.38)). Sensitivity analyses using alternative prompting techniques or the more recent version Claude 3 did not result in substantial changes. Currently, Claude’s RoB 2 judgements cannot replace human RoB assessment. However, the potential of LLMs to support RoB assessment should be further explored.

In small meta-analyses (e.g., up to 20 studies), the best-performing frequentist methods can yield very wide confidence intervals for the meta-analytic mean, as well as biased and imprecise estimates of the heterogeneity. We investigate the frequentist performance of alternative Bayesian methods that use the invariant Jeffreys prior. This prior has the usual Bayesian motivation, but also has a purely frequentist motivation: the resulting posterior modes correspond to the established Firth bias correction of the maximum likelihood estimator. We consider two forms of the Jeffreys prior for random-effects meta-analysis: the previously established “Jeffreys1” prior treats the heterogeneity as a nuisance parameter, whereas the “Jeffreys2” prior treats both the mean and the heterogeneity as estimands of interest. In a large simulation study, we assess the performance of both Jeffreys priors, considering different types of Bayesian estimates and intervals. We assess point and interval estimation for both the mean and the heterogeneity parameters, comparing to the best-performing frequentist methods. For small meta-analyses of binary outcomes, the Jeffreys2 prior may offer advantages over standard frequentist methods for point and interval estimation of the mean parameter. In these cases, Jeffreys2 can substantially improve efficiency while more often showing nominal frequentist coverage. However, for small meta-analyses of continuous outcomes, standard frequentist methods seem to remain the best choices. The best-performing method for estimating the heterogeneity varied according to the heterogeneity itself. Röver & Friede’s R package bayesmeta implements both Jeffreys priors. We also generalize the Jeffreys2 prior to the case of meta-regression.

To estimate accurately the parameters of a regression model, the sample size must be large enough relative to the number of possible predictors for the model. In practice, sufficient data is often lacking, which can lead to overfitting of the model and, as a consequence, unreliable predictions of the outcome of new patients. Pooling data from different data sets collected in different (medical) centers would alleviate this problem, but is often not feasible due to privacy regulation or logistic problems. An alternative route would be to analyze the local data in the centers separately and combine the statistical inference results with the Bayesian Federated Inference (BFI) methodology. The aim of this approach is to compute from the inference results in separate centers what would have been found if the statistical analysis was performed on the combined data. We explain the methodology under homogeneity and heterogeneity across the populations in the separate centers, and give real life examples for better understanding. Excellent performance of the proposed methodology is shown. An R-package to do all the calculations has been developed and is illustrated in this article. The mathematical details are given in the Appendix.

Systematic reviews play important roles but manual efforts can be time-consuming given a growing literature. There is a need to use and evaluate automated strategies to accelerate systematic reviews. Here, we comprehensively tested machine learning (ML) models from classical and deep learning model families. We also assessed the performance of prompt engineering via few-shot learning of GPT-3.5 and GPT-4 large language models (LLMs). We further attempted to understand when ML models can help automate screening. These ML models were applied to actual datasets of systematic reviews in education. Results showed that the performance of classical and deep ML models varied widely across datasets, ranging from 1.2 to 75.6% of work saved at 95% recall. LLM prompt engineering produced similarly wide performance variation. We searched for various indicators of whether and how ML screening can help. We discovered that the separability of clusters of relevant versus irrelevant articles in high-dimensional embedding space can strongly predict whether ML screening can help (overall R = 0.81). This simple and generalizable heuristic applied well across datasets and different ML model families. In conclusion, ML screening performance varies tremendously, but researchers and software developers can consider using our cluster separability heuristic in various ways in an ML-assisted screening pipeline.

Understanding causality is crucial for social scientific research to develop strong theories and inform practice. However, explicit discussion of causality is often lacking in social science literature due to ambiguous causal language. This paper introduces a text mining model fine-tuned to extract causal sentences from full-text social science papers. A dataset of 529 causal and 529 non-causal sentences manually annotated from the Cooperation Databank (CoDa) was curated to train and evaluate the model. Several pre-trained language models (BERT, SciBERT, RoBERTa, LLAMA, and Mistral) were fine-tuned on this dataset and general-purpose causality datasets. Model performance was evaluated on held-out social science and general-purpose test sets. Results showed that fine-tuning transformer models on the social science dataset significantly improved causal sentence extraction, even with limited data, compared to the models fine-tuned only on the general-purpose data. Results indicate the importance of domain-specific fine-tuning and data for accurately capturing causal language in academic writing. This automated causal sentence extraction method enables comprehensive, large-scale analysis of causal claims across the social sciences. By systematically cataloging existing causal statements, this work lays the foundation for further research to uncover the mechanisms underlying social phenomena, inform theory development, and strengthen the methodological rigor of the field.

Effect modification occurs when a covariate alters the relative effectiveness of treatment compared to control. It is widely understood that, when effect modification is present, treatment recommendations may vary by population and by subgroups within the population. Population-adjustment methods are increasingly used to adjust for differences in effect modifiers between study populations and to produce population-adjusted estimates in a relevant target population for decision-making. It is also widely understood that marginal and conditional estimands for non-collapsible effect measures, such as odds ratios or hazard ratios, do not in general coincide even without effect modification. However, the consequences of both non-collapsibility and effect modification together are little-discussed in the literature.

In this article, we set out the definitions of conditional and marginal estimands, illustrate their properties when effect modification is present, and discuss the implications for decision-making. In particular, we show that effect modification can result in conflicting treatment rankings between conditional and marginal estimates. This is because conditional and marginal estimands correspond to different decision questions that are no longer aligned when effect modification is present. For time-to-event outcomes, the presence of covariates implies that marginal hazard ratios are time-varying, and effect modification can cause marginal hazard curves to cross. We conclude with practical recommendations for decision-making in the presence of effect modification, based on pragmatic comparisons of both conditional and marginal estimates in the decision target population. Currently, multilevel network meta-regression is the only population-adjustment method capable of producing both conditional and marginal estimates, in any decision target population.

Results of meta-analyses are of interest not only to researchers but often to policy-makers and other decision-makers (e.g., in education and medicine), and visualizations play an important role in communicating data and statistical evidence to the broader public. Therefore, the potential audience of meta-analytic visualizations is broad. However, the most common meta-analytic visualization – the forest plot – uses non-optimal design principles that do not align with data visualization best practices and relies on statistical knowledge and conventions not likely to be familiar to a broad audience. Previously, the Meta-Analytic Rain Cloud (MARC) plot has been shown to be an effective alternative to a forest plot when communicating the results of a small meta-analysis to education practitioners. However, the original MARC plot design was not well-suited for meta-analyses with large numbers of effect sizes as is common across the social sciences. This paper presents an extension of the MARC plot, intended for effective communication of moderate to large meta-analyses (k = 10, 20, 50, 100 studies). We discuss the design principles of the MARC plot, grounded in the data visualization and cognitive science literature. We then present the methods and results of a randomized survey experiment to evaluate the revised MARC plot in comparison to the original MARC plot, the forest plot, and a bar plot. We find that the revised MARC plot is more effective for communicating moderate to large meta-analyses to non-research audiences, offering a 0.30, 0.34, and 1.07 standard deviation improvement in chart users’ scores compared to the original MARC plot, forest plot, and bar plot, respectively.

Limited evidence on relative effectiveness is common in Health Technology Assessment (HTA), often due to sparse evidence on the population of interest or study-design constraints. When evidence directly relating to the policy decision is limited, the evidence base could be extended to incorporate indirectly related evidence. For instance, a sparse evidence base in children could borrow strength from evidence in adults to improve estimation and reduce uncertainty. In HTA, indirect evidence has typically been either disregarded (‘splitting’; no information-sharing) or included without considering any differences (‘lumping’; full information-sharing). However, sophisticated methods that impose moderate degrees of information-sharing have been proposed. We describe and implement multiple information-sharing methods in a case-study evaluating the effectiveness, cost-effectiveness and value of further research of intravenous immunoglobulin for severe sepsis and septic shock. We also provide metrics to determine the degree of information-sharing. Results indicate that method choice can have significant impact. Across information-sharing models, odds ratio estimates ranged between 0.55 and 0.90 and incremental cost-effectiveness ratios between £16,000–52,000 per quality-adjusted life year gained. The need for a future trial also differed by information-sharing model. Heterogeneity in the indirect evidence should also be carefully considered, as it may significantly impact estimates. We conclude that when indirect evidence is relevant to an assessment of effectiveness, the full range of information-sharing methods should be considered. The final selection should be based on a deliberative process that considers not only the plausibility of the methods’ assumptions but also the imposed degree of information-sharing.

Bibliographic aggregators like OpenAlex and Semantic Scholar offer scope for automated citation searching within systematic review production, promising increased efficiency. This study aimed to evaluate the performance of automated citation searching compared to standard search strategies and examine factors that influence performance. Automated citation searching was simulated on 27 systematic reviews across the OpenAlex and Semantic Scholar databases, across three study areas (health, environmental management and social policy). Performance, measured by recall (proportion of relevant articles identified), precision (proportion of relevant articles identified from all articles identified), and F1–F3 scores (weighted average of recall and precision), was compared to the performance of search strategies originally employed by each systematic review. The associations between systematic review study area, number of included articles, number of seed articles, seed article type, study type inclusion criteria, API choice, and performance was analyzed. Automated citation searching outperformed the reference standard in terms of precision (p < 0.05) and F1 score (p < 0.05) but failed to outperform in terms of recall (p < 0.05) and F3 score (p < 0.05). Study area influenced the performance of automated citation searching, with performance being higher within the field of environmental management compared to social policy. Automated citation searching is best used as a supplementary search strategy in systematic review production where recall is more important that precision, due to inferior recall and F3 score. However, observed outperformance in terms of F1 score and precision suggests that automated citation searching could be helpful in contexts where precision is as important as recall.

When conducting a systematic review, screening the vast body of literature to identify the small set of relevant studies is a labour-intensive and error-prone process. Although there is an increasing number of fully automated tools for screening, their performance is suboptimal and varies substantially across review topic areas. Many of these tools are only trained on small datasets, and most are not tested on a wide range of review topic areas. This study presents two systematic review datasets compiled from more than 8600 systematic reviews and more than 540000 abstracts covering 51 research topic areas in health and medical research. These datasets are the largest of their kinds to date. We demonstrate their utility in training and evaluating language models for title and abstract screening. Our dataset includes detailed metadata of each review, including title, background, objectives and selection criteria. We demonstrated that a small language model trained on this dataset with additional metadata has excellent performance with an average recall above 95% and specificity over 70% across a wide range of review topic areas. Future research can build on our dataset to further improve the performance of fully automated tools for systematic review title and abstract screening.

Systematic searches of published literature are a vital component of systematic reviews. When search strings are not “sensitive,” they may miss many relevant studies limiting, or even biasing, the range of evidence available for synthesis. Concerningly, conducting and reporting evaluations (validations) of the sensitivity of the used search strings is rare, according to our survey of published systematic reviews and protocols. Potential reasons may involve a lack of familiarity or inaccessibility of complex sensitivity evaluation approaches. We first clarify the main concepts and principles of search string evaluation. We then present a simple procedure for estimating a relative recall of a search string. It is based on a pre-defined set of “benchmark” publications. The relative recall, that is, the sensitivity of the search string, is the retrieval overlap between the evaluated search string and a search string that captures only the benchmark publications. If there is little overlap (i.e., low recall or sensitivity), the evaluated search string should be improved to ensure that most of the relevant literature can be captured. The presented benchmarking approach can be applied to one or more online databases or search platforms. It is illustrated by five accessible, hands-on tutorials for commonly used online literature sources. Overall, our work provides an assessment of the current state of search string evaluations in published systematic reviews and protocols. It also paves the way to improve evaluation and reporting practices to make evidence synthesis more transparent and robust.

The number of published systematic reviews has increased over the last years, with a non-negligible proportion displaying methodological concerns. We aimed to develop and evaluate a tool to assess the reported methodological quality of medical systematic reviews. The developed tool (ReMarQ) consists of 26 dichotomous items. We applied an item response theory model to assess the difficulty and discrimination of the items and decision tree models to identify those items more capable of identifying systematic reviews with higher reported methodological quality. ReMarQ was applied to a representative sample of medical systematic reviews (excluding those published in the Cochrane Database of Systematic Reviews) to describe their methodological quality and identify associated factors. We assessed 400 systematic reviews published between 2010 and 2020, of which 196 (49.0%) included meta-analysis. The most discriminative items were (i) conducting a risk of bias assessment, (ii) having a published protocol and (iii) reporting methods for solving disagreements. More recent systematic reviews (adjusted yearly RR=1.03; 95%CI=1.02 −1.04, p<0.001) and those with meta-analysis (adjusted RR=1.34; 95%CI=1.25 −1.43, p<0.001) were associated with higher reported methodological quality. Such an association was not observed with the journal impact factor. The items most frequently fulfilled were (i) reporting search dates, (ii) reporting bibliographic sources and (iii) searching multiple electronic bibliographic databases. ReMarQ, consisting of dichotomous items and whose application does not require subject content expertise, may be important (i) in supporting an efficient quality assessment of systematic reviews and (ii) as the basis of automated processes to support that assessment.

Search filters are single-concept systematic search strategies created by experts. Filters are a valuable resource for systematic searchers. Typically, filters are designed for a single database in a single interface. If researchers do not have access to that specific interface, the existing filter will be unusable without translation. Filter translation is a complex process that requires an understanding of information retrieval concepts, as well as the unique indexing and search functionality of databases and interfaces. The authors undertook a project to translate an APA PsycInfo search filter for Randomized Controlled Trials/Clinical Controlled Trials (RCT/CCT), developed by Canada’s Drug Agency, from the Wolters Kluwer Health Ovid interface to the EBSCO Information Services EBSCOhost interface. We present here a guide for translation, from the first principles of systematic searching to fine details of the relevant database and interfaces, based on our experience and illustrated by a worked example. We discuss each element of a systematic search in a stepwise process, addressing both the underlying information retrieval concepts and the technical strategies for effective translation between the two interfaces. We end with a discussion on translation challenges, with some guidance on how to mitigate potential impacts on sensitivity. While we have endeavored to explain the workings of this process accessibly for researchers who are not experts in systematic searching, anyone undertaking a search translation project should work with a trained information specialist if they lack information retrieval expertise or are unfamiliar with the inner workings of the database, the original interface, and the destination interface.