Research Synthesis Methods最新文献

Several AI-aided screening tools have emerged to tackle the ever-expanding body of literature. These tools employ active learning, where algorithms sort abstracts based on human feedback. However, researchers using these tools face a crucial dilemma: When should they stop screening without knowing the proportion of relevant studies? Although numerous stopping rules have been proposed to guide users in this decision, they have yet to undergo comprehensive evaluation. In this study, we evaluated the performance of three stopping rules: the knee method, a data-driven heuristic, and a prevalence estimation technique. We measured performance via sensitivity, specificity, and screening cost and explored the influence of the prevalence of relevant studies and the choice of the learning algorithm. We curated a dataset of abstract collections from meta-analyses across five psychological research domains. Our findings revealed performance differences between stopping rules regarding all performance measures and variations in the performance of stopping rules across different prevalence ratios. Moreover, despite the relatively minor impact of the learning algorithm, we found that specific combinations of stopping rules and learning algorithms were most effective for certain prevalence ratios of relevant abstracts. Based on these results, we derived practical recommendations for users of AI-aided screening tools. Furthermore, we discuss possible implications and offer suggestions for future research.

While geographic search filters exist, few of them are validated and there are currently none that focus on Germany. We aimed to develop and validate a highly sensitive geographic search filter for MEDLINE (PubMed) that identifies studies about Germany. First, using the relative recall method, we created a gold standard set of studies about Germany, dividing it into ‘development’ and ‘testing’ sets. Next, candidate search terms were identified using (i) term frequency analyses in the ‘development set’ and a random set of MEDLINE records; and (ii) a list of German geographic locations, compiled by our team. Then, we iteratively created the filter, evaluating it against the ‘development’ and ‘testing’ sets. To validate the filter, we conducted a number of case studies (CSs) and a simulation study. For this validation we used systematic reviews (SRs) that had included studies about Germany but did not restrict their search strategy geographically. When applying the filter to the original search strategies of the 17 SRs eligible for CSs, the median precision was 2.64% (interquartile range [IQR]: 1.34%–6.88%) versus 0.16% (IQR: 0.10%–0.49%) without the filter. The median number-needed-to-read (NNR) decreased from 625 (IQR: 211–1042) to 38 (IQR: 15–76). The filter achieved 100% sensitivity in 13 CSs, 85.71% in 2 CSs and 87.50% and 80% in the remaining 2 CSs. In a simulation study, the filter demonstrated an overall sensitivity of 97.19% and NNR of 42. The filter reliably identifies studies about Germany, enhancing screening efficiency and can be applied in evidence syntheses focusing on Germany.

Quantitative evidence synthesis methods aim to combine data from multiple medical trials to infer relative effects of different interventions. A challenge arises when trials report continuous outcomes on different measurement scales. To include all evidence in one coherent analysis, we require methods to “map” the outcomes onto a single scale. This is particularly challenging when trials report aggregate rather than individual data. We are motivated by a meta-analysis of interventions to prevent obesity in children. Trials report aggregate measurements of body mass index (BMI) either expressed as raw values or standardized for age and sex. We develop three methods for mapping between aggregate BMI data using known or estimated relationships between measurements on different scales at the individual level. The first is an analytical method based on the mathematical definitions of z-scores and percentiles. The other two approaches involve sampling individual participant data on which to perform the conversions. One method is a straightforward sampling routine, while the other involves optimization with respect to the reported outcomes. In contrast to the analytical approach, these methods also have wider applicability for mapping between any pair of measurement scales with known or estimable individual-level relationships. We verify and contrast our methods using simulation studies and trials from our data set which report outcomes on multiple scales. We find that all methods recreate mean values with reasonable accuracy, but for standard deviations, optimization outperforms the other methods. However, the optimization method is more likely to underestimate standard deviations and is vulnerable to non-convergence.

RobotReviewer is a tool for automatically assessing the risk of bias in randomized controlled trials, but there is limited evidence of its reliability. We evaluated the agreement between RobotReviewer and humans regarding the risk of bias assessment based on 1955 randomized controlled trials. The risk of bias in these trials was assessed via two different approaches: (1) manually by human reviewers, and (2) automatically by the RobotReviewer. The manual assessment was based on two groups independently, with two additional rounds of verification. The agreement between RobotReviewer and humans was measured via the concordance rate and Cohen's kappa statistics, based on the comparison of binary classification of the risk of bias (low vs. high/unclear) as restricted by RobotReviewer. The concordance rates varied by domain, ranging from 63.07% to 83.32%. Cohen's kappa statistics showed a poor agreement between humans and RobotReviewer for allocation concealment (κ = 0.25, 95% CI: 0.21–0.30), blinding of outcome assessors (κ = 0.27, 95% CI: 0.23–0.31); While moderate for random sequence generation (κ = 0.46, 95% CI: 0.41–0.50) and blinding of participants and personnel (κ = 0.59, 95% CI: 0.55–0.64). The findings demonstrate that there were domain-specific differences in the level of agreement between RobotReviewer and humans. We suggest that it might be a useful auxiliary tool, but the specific manner of its integration as a complementary tool requires further discussion.

In health technology assessment, matching-adjusted indirect comparison (MAIC) is the most common method for pairwise comparisons that control for imbalances in baseline characteristics across trials. One of the primary challenges in MAIC is the need to properly account for the additional uncertainty introduced by the matching process. Limited evidence and guidance are available on variance estimation in MAICs. Therefore, we conducted a comprehensive Monte Carlo simulation study to evaluate the performance of different statistical methods across 108 scenarios. Four general approaches for variance estimation were compared in both anchored and unanchored MAICs of binary and time-to-event outcomes: (1) conventional estimators (CE) using raw weights; (2) CE using weights rescaled to the effective sample size (ESS); (3) robust sandwich estimators; and (4) bootstrapping. Several variants of sandwich estimators and bootstrap methods were tested. Performance was quantified on the basis of empirical coverage probabilities for 95% confidence intervals and variability ratios. Variability was underestimated by CE + raw weights when population overlap was poor or moderate. Despite several theoretical limitations, CE + ESS weights accurately estimated uncertainty across most scenarios. Original implementations of sandwich estimators had a downward bias in MAICs with a small ESS, and finite sample adjustments led to marked improvements. Bootstrapping was unstable if population overlap was poor and the sample size was limited. All methods produced valid coverage probabilities and standard errors in cases of strong population overlap. Our findings indicate that the sample size, population overlap, and outcome type are important considerations for variance estimation in MAICs.

Network meta-analysis (NMA) incorporates all available evidence into a general statistical framework for comparing multiple treatments. Standard NMAs make three major assumptions, namely homogeneity, similarity, and consistency, and violating these assumptions threatens an NMA's validity. In this article, we suggest a graphical approach to assessing these assumptions and distinguishing between qualitative and quantitative versions of these assumptions. In our plot, the absolute effect of each treatment arm is plotted against the level of effect modifiers, and the three assumptions of NMA can then be visually evaluated. We use four hypothetical scenarios to show how violating these assumptions can lead to different consequences and difficulties in interpreting an NMA. We present an example of an NMA evaluating steroid use to treat septic shock patients to demonstrate how to use our graphical approach to assess an NMA's assumptions and how this approach can help with interpreting the results. We also show that all three assumptions of NMA can be summarized as an exchangeability assumption. Finally, we discuss how reporting of NMAs can be improved to increase transparency of the analysis and interpretability of the results.

Sample size and statistical power are important factors to consider when planning a research synthesis. Power analysis methods have been developed for fixed effect or random effects models, but until recently these methods were limited to simple data structures with a single, independent effect per study. Recent work has provided power approximation formulas for meta-analyses involving studies with multiple, dependent effect size estimates, which are common in syntheses of social science research. Prior work focused on developing and validating the approximations but did not address the practice challenges encountered in applying them for purposes of planning a synthesis involving dependent effect sizes. We aim to facilitate the application of these recent developments by providing practical guidance on how to conduct power analysis for planning a meta-analysis of dependent effect sizes and by introducing a new R package, POMADE, designed for this purpose. We present a comprehensive overview of resources for finding information about the study design features and model parameters needed to conduct power analysis, along with detailed worked examples using the POMADE package. For presenting power analysis findings, we emphasize graphical tools that can depict power under a range of plausible assumptions and introduce a novel plot, the traffic light power plot, for conveying the degree of certainty in one's assumptions.

Individual participant data (IPD) meta-analysis projects obtain, harmonise, and synthesise original data from multiple studies. Many IPD meta-analyses of randomised trials are initiated to identify treatment effect modifiers at the individual level, thus requiring statistical modelling of interactions between treatment effect and participant-level covariates. Using a two-stage approach, the interaction is estimated in each trial separately and combined in a meta-analysis. In practice, two complications often arise with continuous outcomes: examining non-linear relationships for continuous covariates and dealing with multiple time-points. We propose a two-stage multivariate IPD meta-analysis approach that summarises non-linear treatment-covariate interaction functions at multiple time-points for continuous outcomes. A set-up phase is required to identify a small set of time-points; relevant knot positions for a spline function, at identical locations in each trial; and a common reference group for each covariate. Crucially, the multivariate approach can include participants or trials with missing outcomes at some time-points. In the first stage, restricted cubic spline functions are fitted and their interaction with each discrete time-point is estimated in each trial separately. In the second stage, the parameter estimates defining these multiple interaction functions are jointly synthesised in a multivariate random-effects meta-analysis model accounting for within-trial and across-trial correlation. These meta-analysis estimates define the summary non-linear interactions at each time-point, which can be displayed graphically alongside confidence intervals. The approach is illustrated using an IPD meta-analysis examining effect modifiers for exercise interventions in osteoarthritis, which shows evidence of non-linear relationships and small gains in precision by analysing all time-points jointly.