Results
We compiled 204 R2 values in total from the collected 56 literature (Tables 1 and S1). Most studies reported the relationships between eDNA quantity and species abundance by the species-specific assay (47 studies) and targeting fish species (40 studies). Contrary, such relationships were reported less frequently for metabarcoding assay (9 studies) and other taxa (7, 4, 3, and 3 studies for herptiles, crustaceans, mollusks, and coral and seastars). The filter pore size and amplicon size ranged from 0 to 10 μm and 52 to 719 bp, respectively. Almost all the studies used filters with less than 3 µm pore size and amplified less than 400 bp target DNA fragments. The number of studies conducted in laboratory, lentic freshwater, lotic freshwater, and marine environments were 18 17, 18, and 6, respectively.
A GLMM with MCMC algorithm showed the significant effects of target taxa, filter pore size, and study environments (Table 2). Relative to fish, estimation accuracy (Fisher’s z-transformed R2values) was significantly lower for crustaceans (posterior mean = -0.544 [95% CI: -1.074, -0.030]) and marginally higher for coral and seastars (0.500 [-0.116, 1.063]). Among the target taxa, the relationship between crustacean eDNA quantity and abundance was the weakest and the estimated R2 values for herptiles and crustaceans were more variable than other taxa (Figure 2a). Laboratory experiments produced higher estimation accuracy than natural environments, whereas there seemed to be no substantial differences in the estimation accuracy among natural environment types (Figure 2b). Filter pore size had a significantly positive effect on the estimation accuracy (0.094 [0,011, 0.180]; Figure 3). In contrast, we did not confirm any statistically significant effects of assay strategy, amplicon size, and abundance metrics (Figures S1 and S2). These statistical trends were not different even if the outliers were excluded (Table S2). We confirmed the convergences of MCMC algorithm for all the run and no substantial multi-collinearity among the factors (adjusted GVIFs: 1.23 to 1.68).
The heterogeneity across studies/datasets was relatively high (I2 = 77.4 % for the overall dataset; Table 3). Although the values were re-calculated by dividing the dataset into each categorical factor (assay strategy, target taxa, study environment, and abundance metrics), the heterogeneities were generally high (> 70 %) except for some categories. In addition, Egger’s regression and funnel plot showed the significant publication bias in our dataset (P < 0.01; Figure 4).