Meta-analyses
We analyzed effect sizes (Hedge’s g ) for both Q1 and Q2 with multi-level meta-analytic (MLMA) models, fitted in R v 4.1.2 (R Core Team 2021) and using the package metafor version 3.0-2 (Viechtbauer 2010). We employed a model selection approach based on the Akaike Information Criterion (AIC) to identify the most important moderators explaining heterogeneity in effect sizes and the most parsimonious model (Arnold 2010). This required first fitting the full model and all reduced models via maximum likelihood (ML) estimation. For Q1, the full model included the moderator variables infection status, fitness trait, stressor type, and all their interactions. The full model for Q2 included response trait, stressor type, and their interaction.
All models accounted for non-independence of effects and sampling errors measured in the experiment. Models also included observation-level random intercepts, so residual variation within studies could be estimated. Full and reduced models (including intercept-only model) were compared using the ‘dredge’ function of the R package MuMIn v 1.43.17 (Bartón 2020). The highest-ranking model based on small sample size corrected AIC (AICc) was then refitted via restricted maximum-likelihood (REML) estimation to interpret moderators and evaluate publication bias and heterogeneity.
We report meta-analytic mean estimates and 95% confidence intervals for effects of moderators in final models. Meta-analysis results were plotted using the R package orchaRd (Nakagawa et al.2021). We tested significance of statistical contrasts between fitness and infectivity response variables in Q2 using Wald-type chi-square tests, computed with the function ‘anova’.