Rapid warming could drastically alter host-parasite relationships, which is especially important for fisheries crucial to human nutrition and economic livelihoods, yet we lack a synthetic understanding of how warming influences parasite-induced mortality in these systems. We conducted a meta-analysis using 266 effect sizes from 52 empirical papers on harvested aquatic species and determined the relationship between parasite-induced host mortality and temperature and how this relationship was altered by host, parasite, and study design traits. Overall, higher temperatures increased parasite-induced host mortality; however, the magnitude of this relationship varied. Hosts from the order Salmoniformes experienced a greater increase in parasite-induced mortality with temperature than the average response to temperature across fish orders. Opportunistic parasites were associated with a greater increase in infected host mortality with temperature than the average across parasite strategies, while bacterial parasites were associated with lower infected host mortality as temperature increased than the average across parasite types. Thus, parasites will generally increase host mortality as the environment warms; however, this effect will vary among systems.

We compiled data from experimental studies on fisheries species that compared mortality of parasitized and unparasitized hosts at a static temperature. We defined fisheries species to include both invertebrate and vertebrate species that are harvested commercially or recreationally. In Fall 2019, we searched Web of Science following PRISMA protocols (Foo et al. 2021) using key terms that would return papers focused on harvested aquatic species, parasites, and diseases, but would exclude papers that were focused on human, environmental or domestic animal health (see Appendix S1 in Supporting Information). This search yielded 1,201 papers. We then screened the abstracts of these papers, and retained only papers that satisfied four criteria: 1) an experiment was performed that included at least one parasite exposure treatment paired with an unexposed control group, 2) temperatures were intended to be constant and not intentionally varied, 3) hosts were from species that constitute a fishery, including those in aquaculture, and 4) estimates of survival or mortality were reported for both infected and uninfected hosts at each temperature treatment.  This selection process reduced the number of studies to 386 (Appendix S1 and Figure S1). 

We then screened the abstracts of these papers, and retained only papers that satisfied four criteria: 1) an experiment was performed that included at least one parasite exposure treatment paired with an unexposed control group, 2) temperatures were intended to be constant within a temperature treatment and not intentionally varied, 3) hosts were from species that constitute a fishery, including those in aquaculture, and 4) estimates of survival or mortality were reported for both infected and uninfected hosts at each temperature treatment.  This selection process reduced the number of papers to 386 (Appendix S1 and Figure S1). 

We obtained full versions of 358 papers (28 papers from the original 386 were unobtainable).  We then screened the full text of these papers to ensure a match to our four criteria, which reduced the 358 papers to 70. To increase statistical power to estimate the effect of host Order on parasite-induced mortality, we excluded experiments from hosts in Orders with fewer than two experiments with ≥ 2 temperature treatments. We also removed experiments with zero mortality in infected and uninfected treatments at all experimental temperatures. These experiments provided no information about how parasite-induced mortality changed with temperature. This reduced the number of papers included in our dataset from 70 to 52 and yielded a total of 266 effect sizes from 121 experiments (several papers included more than one experiment; Appendix S1 and S2, Figure S1). 

At least two people extracted data from each paper to reduce extraction error. If extracted values differed, the data were re-extracted until there was agreement between the two extractors. For data that were displayed in a graphical format only, we used WebPlotDigitizer (Rohatgi 2022) to extract data. Data (which may have been presented as mortality rates, or proportion surviving) were converted to numbers of host individuals that were dead and alive at the end of the experiment. We also extracted information about the paper itself, including the source of the hosts used in the paper and the motivation for conducting the experiment (see Appendix S1). Finally, we collected additional information about host and parasite traits from outside sources (e.g., other peer reviewed papers, government reports) when necessary to obtain moderator variables (Table 1, Appendix S1 and S2). The moderators (Table 1) were used to test a priori hypotheses regarding how host, parasite, and study design traits influenced how temperature affected parasite-induced mortality.

Because our focus was on parasite-induced mortality, we used log odds ratios and the variance surrouding log odds ratio as our effect size to compare host survival in the parasitized vs unparasitized treatments.

Similar to Sauer et al. (2020), we used a mixed effect meta-analysis (metafor package, rma.mv function; Viechtbauer 2010) to analyze our data. This model accounts for heterogeneity among experiments by allowing the relationship between temperature and LOR to have a different intercept for each experiment. The overall slope provides an estimate of how LOR changes with temperature on average.  Additionally, the random intercepts account for variation in the relationship between LOR and temperature that results from factors we were unable to model explicitly (e.g., dosage, geographic source of host or parasite, laboratory facilities).This model accounts for heterogeneity among experiments by allowing the relationship between temperature and LOR to have a different intercept for each experiment. The overall slope provides an estimate of how LOR changes with temperature on average.  Additionally, the random intercepts  account for variation in the relationship between LOR and temperature that results from factors we were unable to model explicitly (e.g., dosage, geographic source of host or parasite, laboratory facilities).

We explored twelve versions of equation 3 that avoided collinearity that arose between host taxonomic information and parasite or host traits. In addition to running the above models on the full dataset, we also applied two trait-based models, single categorical trait models, and a null model on a reduced dataset that was restricted to the Order Salmoniformes, the most common Order in the data set (comprising 26 % of the effect sizes). This allowed us to explore temperature effects in a well-studied, and economically important, group. 

Excluding the null models, we used backwards elimination to produce a model that included only factors with significant interactions with temperature (p < 0.1) in both the full and Salmoniformes datasets.  All the models generated in this process were compared using an Information-Theoretic approach that utilizes Akaike Information Criterion corrected for sample size (AICc; Akaike 1973; Blankenship et al. 2002; Whittingham et al. 2006) to select the best fit model. We also calculated McFadden’s Pseudo–R² (Hemmert et al. 2018) and Nakagawa’s Pseudo-R² for generalized linear mixed-effects models (orchaRd package, r2_ml function; Nakagawa et al. 2023; Nakagawa & Schielzeth 2013) which estimates the amount of additional variation explained by the mixed-effects model compared to the null model. Nakagawa’s Pseudo-R² provides a goodness-of-fit metric for both the fixed-effects of the model (Marginal R²) and for the full mixed-effects model (Conditional R²). While our models estimate both main effects and interactions with temperature of our fixed effects, we discuss only the interactions because these are the estimates that reflect how the effects of temperature on parasite-induced mortality are influenced by host, parasite, and study design traits.

We additionally evaluated publication bias and the degree of variance caused by heterogeneity in the true effect sizes compared to the variance caused by sampling error in our data. To evaluate publication bias we assess the effect of low sample size studies on the overall effect and if the overall effect declines with publication year using the procedures outlined in Nakagawa et al. 2022 (Appendix S3) for both the full and Salmoniformes datasets. We calculated the typical sampling error variance from studies contained in our meta-analysis to estimate the statistical noise in our data (sigma2_v function; Yang et al. 2023) as well as the I² metric to interpret the heterogeneity in our data and determine the variability in the true effect attributed to each level of our random effects (Higgins & Thompson 2002; orchaRd package, i2_ml funciton, Nakagawa et al. 2023.) All analyses were performed using R version 4.4.2 (R Core Team 2024).