Data collection and transformations
We obtained primary literature data directly from main text, tables,
supporting material, or raw data files whenever available. Otherwise, we
digitized data from figures using PlotDigitizer
(https://plotdigitizer.com). Stressor effects were standardized to
unbiased mean differences (Hedge’s g ) from both continuous and
discrete variables (Hedges 1981). For continuous variables, we obtained
mean and standard deviation (SD) of fitness traits and infectivity
metrics in environments with different exposure to stressors. If SD was
not reported, an error estimate (standard error (SE), 95% confidence
interval (CI) or Wald’s CI) was converted to SD, assuming normality. If
a study reported median instead of mean (n = 13 effects in four
studies), we estimated the mean following Hozo et al. (2005). If
dispersion was only reported as data range or interquartile range (n = 8
effects in one study and n = 5 effects in three studies, respectively),
we approximated SD (Lajeunesse 2013; Wan et al. 2014). Mean and
SD of response variables were then used to calculate standardized mean
differences (d) and their variances.
Many studies (n = 67) used discrete variables to quantify infection
prevalence and/or survivorship. In these cases, we calculated odds
ratios between environmental treatments and estimated variances
(Rosenberg et al. 2013). In cases where at least one category had
no observations (e.g., no survival in polluted treatment), we applied
Yate’s continuity correction to avoid dividing by zero (Yates 1934). Log
odds ratios were then converted to d , and variances of log odds
ratios were converted to variances of d , assuming a continuous
logistic distribution underlies each discrete trait (Hasselblad &
Hedges 1995). Finally, we estimated Hedge’s g and its variance by
applying sample size correction J to all values of d and
their variances (Hedges 1981).
Most experiments (n = 108) contrasted host fitness traits and
infectivity across three or more environmental treatments or in more
than one-time interval. For example, a control group could be compared
to two levels of chemical pollution or at both 24 and 48 hpi. In these
cases, stressor effects and sampling errors were not independent, as
they shared control group or time baseline. To account for correlated
sampling errors between these effects, we computed covariances in
sampling errors between effects in multiple-comparison designs following
Viechtbauer (2010). We included these variance-covariance matrices in
our statistical analyses (see below). For a few experiments (n = 8)
where large covariances between effects and small sample sizes resulted
in variance-covariance matrices with negative eigenvalues (i.e., not
positive definite), we adjusted covariance estimates to produce the
nearest positive definite matrix using the R package Matrix(Douglas & Maechler 2021). As an alternative approach to estimating
sampling error covariances, we adjusted fixed effect coefficients using
the robust variance estimator (RVE) (Hedges et al. 2010), as
implemented in the R package clubSandwich (Pustejovsky 2020).
Here, we focus on results with estimated covariances and show results
under the RVE in Supporting Material.