Model selection
We used R (v3.5.3; R Core Team, 2019, Vienna Austria) for all analyses. We constructed either linear mixed effects models using the lme4 package or linear models using stats package to test our four main hypotheses. Final best-fit models were selected using an AIC selection criterion, ΔAIC ≥ 3. The only exception was when ΔAIC was within 3 for the models for liver and pectoralis lipid damage including cohort as a random factor compared with excluding it, and in these cases, we decided to control for cohort.