2.4 Statistical analysis
All statistical analyses were performed with R (version 3.4.1). The different statistical models built to analyze the data are described in the supplementary material (Table S1 ). Mortality during development (death = 0, emerged = 1) was analyzed with Generalized Linear Model (GLM) with binomial distribution of errors. Explanatory variables were glyphosate concentration and glyphosate type (pure or in formulation) for the experiment 1 and glyphosate exposition and food treatments for the experiment 2. A GLM with normal distribution of errors was used to test for difference in development time (calculated as the number of days from hatching to emergence) and wing length (mm) among the larval treatments. Explanatory variables were mosquito sex, glyphosate concentration and glyphosate type (pure or in formulation) for the experiment 1 and mosquito sex, glyphosate exposition and food treatments for the experiment 2. Female mosquito-centred traits (blood meal size, number of eggs, infection prevalence and oocyst burden), which may depend on which bird mosquitoes fed on, were analyzed fitting bird as a random factor into the models (to account for the spatial pseudoreplication), using lmer or glmer (package: lme4, Bates et al., 2015) according to whether the errors were normally (blood meal size, number of eggs and oocyst burden) or binomially (prevalence) distributed. Glyphosate concentration and glyphosate type (pure or in formulation) for the experiment 1 and glyphosate exposition and food treatments for the experiment 2 were used as fixed factors. Blood meal size was also added as a fixed factor when it was not a response variable.
Maximal models, including all higher-order interactions, were simplified by sequentially eliminating non-significant interactions and terms to establish a minimal model (Crawley, 2012). The significance of the explanatory variables was established using either a likelihood ratio test (which is approximately distributed as a Chi-square distribution, Bolker, 2008) or an F test. The significant Chi-square or F values given in the text are for the minimal model, whereas non-significant values correspond to those obtained before the deletion of the variable from the model. A posteriori contrasts were carried out by aggregating factor levels together and by testing the fit of the simplified model using LRT (Crawley, 2012).
Results