Statistical Analyses
Statistical analyses were conducted in R version 4.2.0 (R Core Team 2022). All data were checked for normality and homoscedasticity. To determine the effect of diet treatment and time (day 0, day 8) on body mass, fat score, corticosterone concentrations, and complement (CH50) activity in experiment 1, we ran GLMMs using the glmmTMB package in R (Brooks et al. 2017). For body mass, we ran a GLMM fitted with a gaussian distribution, while the models for the remaining endpoints (fat score, corticosterone, and complement activity) were fitted with a tweedie distribution. All models included treatment, time (day 0, day 8), and the interaction between treatment and time as fixed effects with bird ID as a random effect. When testing the effect of diet treatment and time on corticosterone responses, we also included sample type (baseline or stress-induced) as a fixed effect in the model. To test whether the average grams consumed per day varied based on diet treatment, we ran a generalized linear model (GLM) fitted with a gaussian distribution with grams consumed as the response variable and treatment as a fixed effect. Because physiology can both mediate and respond to shifts in gut microbiota composition (Noguera et al. 2018, Williams et al. 2020, Yoshiya et al. 2011) and diet might influence these interactions, we also ran a separate GLMM for each alpha diversity metric (observed richness, Shannon diversity) which considered the interactions between diet treatment and time (day 0, day 8) with each physiology metric (complement activity, baseline corticosterone, stress-induced corticosterone). P-values from these models were adjusted for false discovery rate with a Benjamini–Hochberg correction where significance was determined as Padj < 0.05. We used the DHARMa package (Hartig 2019) to plot residuals and confirm the suitability of each model and the Anova function in thecar package (Fox and Weisberg 2018) to determine significance.
To examine how feeding behavior and diet preference changed over time with treatment in experiment 2 (diet selection), we ran separate GLMs fitted with a gaussian distribution to test for differences in the stimulus (LPS, saline) and focal (LPS-focal, saline-focal) groups, respectively. Because birds were housed in pairs for this experiment (Figure 1), pre- and post-treatment feeding behavior was analyzed by cage rather than by individual. To examine how baseline corticosterone concentrations and hemolytic complement activity varied in the focal birds, we ran GLMMs for each response variable where we included treatment (LPS-focal or saline-focal), time, an interaction between treatment and time, and sex as predictors. Testosterone samples were collected from males only, so sex was not included in this model. For body mass and fat score, we ran a GLMM fitted with a gaussian distribution, while the models for the remaining endpoints (baseline corticosterone, complement activity, and testosterone) were fitted with a tweedie distribution. To investigate whether differences in feeding behavior affected the gut microbiota, we also ran a separate GLMM for each alpha diversity metric (observed richness, Shannon diversity), which considered the interactions between treatment and time (day 0, day 5) as well as the average amount of grams of each diet type (fat diet, protein diet) consumed per day pre- and post- treatment. P-values from these models were adjusted for false discovery rate with a Benjamini–Hochberg correction where significance was determined as Padj < 0.05. Due to power limitations associated with low sample sizes for focal birds that had both physiological and gut microbiota samples for days 0 and 5 (n=2-9 depending on the physiological metric), we were unable investigate interactions between physiology (complement, baseline corticosterone, testosterone), treatment, and time for gut microbial metrics as we did in the first experiment.