2.4 Multistate mark-recapture models
To estimate the probabilities of survival (Φ), encounter (or recapture, ρ) , and transition (ψ ) between states (breeder, non-breeder, and prebreeder), multistate mark-recapture models were constructed using program MARK and the ‘RMark’ package in R (White & Burnham 1999; Laake 2013). Within these models, a group effect for colony (Robben Island and Stony Point) was included to evaluate colony differences in the estimates. Known parameters were fixed to improve model performance; since only breeders were marked in 2013 across both colonies (with nonbreeders marked in subsequent years, Table A1), survival and transition rates for nonbreeders and prebreeders were fixed to zero during 2013–14, as was recapture in 2014 in both colonies. Additionally, no nonbreeders were marked in 2014 at Robben Island, so prebreeder survival and transition during 2014–15, and prebreeder recapture in 2015 were also fixed to zero for this colony.
Initially, a general model was developed assuming time, state, and colony dependence for survival, recapture, and transition probabilities. Simpler model structures were also tested for recapture whereby years were pooled into two groups to represent before and after ground readers were installed in each colony. For survival and transition, simpler models were also included whereby time dependence was replaced with both combined and separated annual sardine and anchovy spawner biomass WoCA to determine whether fish abundance offered better predictive power for survival and transition probabilities than the fully time-dependent model.
Recapture probabilities were modelled first, with the best fitting model taken forward to assess survival probabilities, followed by transition. Model selection was based on Akaike’s Information Criterion corrected for over-dispersion and small sample size (QAICc) (Lebreton et al. 1992). When models differed by QAICc <2, they were considered approximately equivalent (Burnham & Anderson 2002), and the model with the lowest number of parameters was considered the most parsimonious. Goodness-of-fit (GOF) tests for the general model (JMV model; Pradel et al. 2003) were performed using package ‘R2ucare’ (Gimenez et al. 2018) in Program R.