Diversity metrics and meta-analysis

We computed individual-based diversity metrics using the function hilldiv::hill_div, and obtained average alpha diversity metrics per species, as well as wild and captive populations per species. We used a Kruskal–Wallis (KW) test as implemented in the function hilldiv::div_test to ascertain whether the mean diversity values varied across analysed host species, and a PERMANOVA (PMV) test using vegan::adonis function based on the pairwise dissimilarity matrix to test whether host species were compositionally distinct.

Average alpha diversity metrics of wild and captive populations per species were used to conduct a random-effects-model (REM) meta-analysis with raw effect sizes using the function meta::metacont. We used the Sidik–Jonkman estimator for the between-study variance and the Knapp–Hartung–Sidik–Jonkman adjustment method. The overall effect was calculated using Hedge's g (SMD) and its 95% confidence interval and p-value. An identical analysis was performed for the entire dataset and two representative subsets of five species, containing only datasets derived from primates and cetartiodactylans. Higgin’s and Thompson’s I² test, Tau-squared T² and Cochran’s Q were used for quantifying the heterogeneity between the included species. Due to the high heterogeneity found in the study, we evaluated whether the between-study heterogeneity was caused by outliers with extreme effect sizes, which could be distorting our overall effect. We defined an outlier if the species’s confidence interval did not overlap with the confidence interval of the pooled effect using dmetar::find.outliers function.The function detected three outliers in dR metric (GOGO, PEMA and TUTR), four in dRE (GOGO, PEMA, MOCH, EQKI) and seven in dRER (RHBR, PYNE, PEMA, TUTR, MOCH, CENI and AIME). Even when these outliers were excluded from the analysis the I² heterogeneity value was substantial for dR (I² from 79.3 to 70.3%) and moderate for dRE (I² from 80.1 to 60.0%) and dRER (I² from 86.9 to 54.2%) and significant for both (Cochran’s Q, p-value < 0.001). We performed a sensitivity analysis removing the outliers from the meta-analysis, yet the results of the random effects model did not change (dR: SMD = 0.345, p-value = 0.075; dRE: SMD = 0.021, p-value = 0.901; dRER: SMD = 0.015, p-value = 0.928). We also performed Graphical Display of Study Heterogeneity (GOSH) plots to explore the patterns of effect sizes and heterogeneity in our data. We used three supervised machine learning (k-means, DBSCAN and the Gaussian Mixture Model) algorithms to detect clusters in the GOSH plot data and determine which studies contribute the most to them automatically using dmetar::gosh.diagnostics function.