orchaRd 2.0: An R package for visualising meta‐analyses with orchard plots

Meta-analysis has become an essential synthesis tool across the medical, social and biological sciences (Cooper et al., 2019; Gurevitch 2018; Higgins Schmid 2021). In fields such as medicine, meta-analytic results are typically shown in a forest plot that presents effect sizes their 95% confidence intervals (CIs) from each study meta-analysis. However, ecology evolution, plots infrequently used because meta-analyses this field often include >100 sizes, making traditional impractical (Gurevitch Senior 2016). Instead, researchers use ‘forest-like plot’ with overall mean size estimate CIs for different levels of categorical moderator (predictor variable). Such estimates derived meta-regression model or subset/sub-group analyses. For example, could show five taxa, six geographical areas three methods. A recent survey found 72 out 102 ecological evolutionary presented forest-like (Nakagawa Contributing to popularity is fact moderators rather than continuous variables. Despite popularity, evolution lack important information individual heterogeneity among 2021; Schild & Voracek, 2015). Nakagawa al. (2021) introduced information-rich version plot, named ‘orchard’ plot. Orchard provide (1) point (i.e. means); (2) CIs, (3) prediction (PIs; which sizes); (4) scaled by precision (the inverse square root sampling variance). implemented orchard using functions most popular comprehensive meta-analysis R package, metafor (Viechtbauer, 2010) ggplot2 graphics (Wickham, 2009). original implementation (orchaRd 1.0) was limited single models. addition, it only possible visualise assumed homoscedasticity all have same variance, may be unrealistic, e.g. Wilson 2022; Zajitschek 2020). article, we enhance visualisation capabilities orchaRd package integrating functionalities emmeans (Lenth 2018) four ways. The first extend allowing (I) heteroscedasticity (different variances moderator; Section 3.1); (II) marginal (e.g. marginalising other apart one visualised; 3.2) (III) conditional groups/levels variable, conditioned upon specific values variable; 3.3). fourth capability allows ‘bubble’ created (i) (ii) interactions between variable (iii) two variables multi-moderator models (Section 3.4). add helper calculate key statistics multilevel I2 (Cheung, 2014) R2 (Aloe 2010; Schielzeth, 2013), along 3.5). These new not better but also facilitate exploration previously neglected patterns, data. Throughout support motivation creating these 2). Notably, 2.0 improves reporting transparency following ‘Preferred Reporting Items Systematic reviews Meta-Analyses Ecology Evolution’ (PRISMA-EcoEvo; O'Dea Our package's vignette provides detailed instructions examples on how main functions, well customise (https://daniel1noble.github.io/orchaRd/). To gauge potential usefulness extensions, surveyed evolution. dataset initially collected quantify quality assist PRISMA-EcoEvo (O'Dea Briefly, obtained articles were published 1 January 2010 25 March 2019 part ‘Ecology’ ‘Evolutionary Biology’ journals classified under InCites Journal Citation Reports (Clarivate Analytics; see more details We explored report relevant below, full can Supporting Information. (‘table’, ‘figure’ ‘statistics’) functionalities: mod_results (creating table function; Figure 1), orchard_plot (a figure function), bubble_plot caterpillars (5) i2_ml (calculating function) (6) r2_ml function's description Table 1). Among core function orchard_plot. This enables users draw mod_results, uses functionality process objects (object classes: rma, rma.mv robust.rma; Viechtbauer, 2010). Below showcase Then, describe function, bubble_plot, followed (caterpillars, r2_ml). focus our models, deal types non-independence due correlated multiple per study) errors shared control groups measurements; Noble 2017). former requires adding random effects ID), while latter modelling within-study variance–covariance matrix (note vcalc create matrix; Categorical (moderators) extremely common meta-analyses. survey, >97% papers had at least variable. subset data sub-group analyses, where series (intercept models) run, fit (Q1). many meta-analyses, variation (homoscedasticity). shows 5% investigated heteroscedasticity, others (Q2). Yet, differences biologically insightful means groups. Pottier (2022) aquatic ectotherms thermally plastic terrestrial counterparts, responses much those (even after considering sample difference). now modelled depicting PIs (Figure Of importance, when exists, might reduce Type error (Rubio-Aparicio 2017, 2020); meta-analysts finding heteroscedasticity. Incidentally, becomes if wants obtain absolute group selection gradients; Kingsolver 2012; Siepielski 2017; 2018). Absolute calculated assuming ‘folded’ normal distribution (see Morrissey, 2016; Lagisz, 2016), accuracy magnitudes being dependent within-group variances. As such, evaluated approach taken. Many (moderators), they together model. studies (Q3: 41%) (Q4: 30%). Not reported (Q5: 27%). It understandable obtaining ‘marginal’ difficult once number increases unless relies computational solutions, via package. Therefore, been uni-moderator made straightforward produce notable marginalisation usually done weighting proportion frequencies (data) averaged over. case, similar, identical, ‘equal’ (giving weights groups), marginalised model, especially unbalanced Equal is, useful your unequal dataset, population, should ~50:50%; males females animals (cf. Deffner 2022). mentioned above, showed uncommon combination, group-level (and means) 3). (2019) estimated thermal environments during development affect phenotypic variance. They increasing temperature did change phonotypic means, variance increased developmental increased. Examining ‘conditional’ illuminating statistical inference significance gradient moderator. none 32 containing estimates, example depicted 3b Vendl ~30 (out 102) some type interaction (Q5). Three manifest meta-analysis, categorical–categorical variables; categorical–continuous continuous–continuous (categorical–categorical) easily visualised conceptualised 2 another equivalent 4 levels; 4a). If want second (categorical–continuous), bubble plots, called regplot, single-panel unlike multi-panel plots; 4b). third (continuous–continuous) intuitive visualise, ‘bubbleless’ line panels 4c); bubbleless there few no corresponding points given addition ‘caterpillar’ (via caterpillars, without labels size; vignette—https://daniel1noble.github.io/orchaRd/). present non-plot give convenient tools explained calculates I2, percentage driven (much studies; Thompson, 2002). (referred ‘total’ I2) additional ID species ID; sensu Santos, 2012). Furthermore, sets (levels) While (intercept-only) (heterogeneity) accounted moderators. R2, proposed Schielzeth (2013) pseudo-R2 linear mixed-effects both bootstrapping. recommend systematic Visualisations completely consistent recommendations PRISMA-EcoEvo. so (sub-)items, recommended Method section: presenting numbers estimate; indicators heterogeneity; including conducted very poor items: 57%, 52% 59%, respectively. see, takes care items (Figures 1-4). even heterogeneities heteroscedasticity) PIs. Graphical presentation (3.1%) 64 R, cited any graphical package(s) visualising orchaRd; Q8). starkly contrasts 85% (55 64; Q7) citing software packages metafor). result marks severe under-recognition packages. real-world risk here recognition severely disincentivises developers maintaining further developing argue authors acknowledge (or research matter), just do propose make listed end legend. standardised format will necessarily need methods, still credit. note, however, dependencies required ‘base’ packages). emmeans, metafor. freely admit satisfying answer whether credited. think reasonable suggest reference (in legend text) immediate figure. presence influence grow field, ever manner. Here, expanded (version 2.0), readily complex simple results, task many. New allow plotted improve communication holistic visual interpretation numerical generated analysis Also, introduce calculating standard Finally, hope paper reminder importance acknowledging Adequate attribution credits sustainable environment maintainers Shinichi Daniel W. A. conceived initial idea wrote draft. led programming implementations inputs Rose E. O'Dea, Alistair M. Patrice Pottier. Malgorzata Joanna Rutkowska Yefeng Yang survey. All contributed design editing commenting drafts. thank Wolfgang Viechtbauer his effort maintain develop amazing Russell Lenth wonderful compatible Lagisz supported ARC (Australian Research Council) Discovery Grant (DP210100812), (DP210101152). Part writing visiting Okinawa Institute Science Technology (OIST) through Theoretical Sciences Visiting Program (TSVP) Nakagawa. Open access publishing facilitated University South Wales, Wiley - Wales agreement Council Australian Librarians. author conflict interest. peer review history article available https://www.webofscience.com/api/gateway/wos/peer-review/10.1111/2041-210X.14152. code GitHub (https://github.com/daniel1noble/orchaRd) Zenodo (https://doi.org/10.5281/zenodo.7928743; 2023). Please note: publisher responsible content supporting supplied authors. Any queries (other missing content) directed article.