General modelling procedures
We fitted linear mixed-effect models using a Bayesian framework
implemented in R v3.3 (R Core Team 2018) with the package MCMCglmm
(Hadfield 2010). We ran 1,100,000 iterations per model, from which we
discarded the initial 100,000 (burn-in period). Each chain was sampled
at an interval of 500 iterations, which resulted in low autocorrelation
(<0.05) among thinned samples. Posterior modes, 95% credible
intervals (CI) and (co)variances were estimated across the thinned
samples for the fixed and random effects. Fixed-effect priors were
normally distributed and diffuse with a mean of zero and a large
variance (100). We explored the sensitivity of the variance-covariance
matrix to the choice of prior. See Appendix S1 for prior details on each
of the analyses. Mean values of the posterior distributions were robust
to different relatively uninformative priors. However, the width and
mode of the posterior distribution for the animal model was susceptible
to prior choice. We thus decided to present the animal model results
estimated with restricted maximum likelihood framework using the package
ASreml-R v.4.
Results