Statistical analyses
All the statistical analyses were performed by using R version 4.0.4 (R Core Team, 2021). Among the collected literature, we excluded some of them as a matter of convenience in the further analysis (Appendix S1). To test the categorical effects (taxa and environment) on estimation accuracy of species abundance, we estimated the effect sizes and their variances by Fisher’s z‐transformation, which allows to avoid ‘Simpson’s paradox’ and compare R2 values among categories accounting for sample sizes. Using the metacor function in the package ‘meta’ (Balduzzi et al., 2019), we produced forest plots to integrate R2 values with their 95% confidence intervals (CIs) among multiple individual datasets, where weighted-average R2 values were calculated by using the inverses of variances estimated above, using the rma function in the package ‘metafor’ (Viechtbauer, 2010). We adopted the random-effect model, assuming that all datasets share a common effect size but also vary among datasets, as our collected dataset is not functionally identical (i.e., some R2 values could be derived from the same study). More strictly, the weight (inverse variance) for averaging R2 values comprised both intra- and inter-study variances of effect size (Borenstein et al., 2010).
In addition, to test the quantitative effects (filter pore size and PCR amplicon size) on estimation accuracy of species abundance, we performed a linear mixed modeling (LMM) by the lmer function in the package ‘lmerTest’ (Kuznetsova et al., 2017). Prior to model fitting, we logit-converted R2 values to meet normality and included the values as the dependent variable. Filter pore size (µm) and amplicon size (bp) were explanatory variables. As random effects, we included study groups, abundance metrics (biomass/density), target taxa, and environment, assuming that the effects of filter pore size and amplicon size could be underestimated without consideration of these categorical valuables. Because almost all studies included in the model targeted shorter eDNA fragments (<200 bp), we excluded the dataset targeting 719-bp fragments of mitochondrial eDNA (Jo et al., 2017) as an outlier.