1 Appendix S1: Bayesian methods for parameter estimation 2 3 Bayesian methods have become popular for analyzing mark-recapture data. Since 4 the early studies (see Clark et al. 2005 for a literature review), methods have been widely 5 applied and customized for specific applications (e.g. Royle & Dorazio 2008). The trends 6 for methodological development have included using hierarchical structure to model 7 heterogeneity, dealing efficiently with missing data, and using covariates to model 8 parameters of interest. We use Bayesian methods for estimating parameters for multistate 9 mark-recapture data (Calvert et al. 2009, Clark et al. 2005). Other authors have used 10 Bayesian methods to estimate movement in open population models (Dupuis & Schwarz 11 2007), examine age-dependence in survival (Zheng et al. 2007), model metapopulation 12 dynamics (Royle & Kery 2007), model individual effects in survival models (Royle 13 2008), and estimate population size using data augmentation for both closed and open 14 populations (Royle & Dorazio 2010). 15 We developed a software program in MATLAB that implements a Bayesian 16 technique for estimating population vital rates from mark-recapture data with multiple 17 life stages. The Bayesian framework for mark-recapture data allows the probabilistic 18 estimation of vital rate parameters (survival, transition and recapture probabilities). 19 Instead of point estimates, the Bayesian approach gives the solution as a distribution of 20 different possible vital rates that explain the observed data. These distributions are often 21 analytically intractable, but there are numerical methods to approximate the distributions 22 by producing random samples from them. We used a Gibbs sampling algorithm for 23 estimating population vital rates, described in detail here. 1 1 We estimated separate survival, transition and recapture probabilities for each 2 year and each life stage. This model is developed for structured populations, where the 3 population can be stratified by life stage at each time step (Dupuis 1995). We used the 4 following notation: 5 xkt observed life stage for individual k at year t 6 zkt true life stage for individual k at year t 7 Sit survival probability from year t to year t+1 for life stage i 8 Rit recapture probability in year t for life stage i 9 Pijt transition probability from life stage j at year t to life stage i at year t+1 10 There are four life stages i, represented by the following numbers: 1=juvenile, 11 2=subadult, 3=male, and 4=female. 12 Raw mark-recapture data were entered directly into the model, where two 13 matrices summarized the capture history for a given individual (Dupuis 1995). Each entry 14 in matrix x was the observed life stage of an individual. If the individual was not 15 observed, we set xkt = 0. Following Clark et al. (2005), we defined the matrix z, which 16 contained the estimated life stage of an individual (which includes both the observed and 17 unobserved life stages). The first matrix consists of raw data, and life stage values for the 18 other matrix are unknown state parameters, estimated as part of the Bayesian analysis. 19 From a previous analysis, we determined that capture probabilities varied by both 20 life stage and year (McCaffery & Maxell 2010). We used the secondary sampling 21 sessions within each summer to estimate the recapture probability for that year. Thus, the 22 likelihood for observing x for a given year t, with recapture probabilities R was 4 P(x*t | R*t ) Rit oit (1 Rit ) uit 23 (1) i1 2 1 where the indicator variables okt and ukt indicate the number of observed and unobserved 2 individuals at year t and life stage i, respectively (counted from all secondary capture 3 sessions). Survival and transition probabilities were estimated between primary capture 4 sessions. The likelihood for life stages z for a given year t, with survival and transition 5 probabilities S and P was 4 P(z*t | S*t ) Sit Sit (1 Sit ) d it 6 (2) i1 4 7 4 P(z*t | P**t ) Pijt #(i j )t (3) i1 j 1 8 where indicator variables sit, dit, and #(ij)t indicate the number of survivals, deaths and 9 transitions at year t, respectively (counted between primary sessions). 10 For estimating the true life stages, we followed Clark et al. (2005) and 11 conditioned the life stage at time t on life stages at the previous and next time point 12 (surviving previous time, stage transition, not capturing, surviving next time and stage 13 transition): 14 15 P(zkt j | zk,t 1 m,zk,t 1 l,R,S,P) Sm,t 1Pjm,t 1 (1 R jt )S jt Pljt (4) The true life stages z are implicitly connected to the observed life stages x: the 16 value of x affects the neighboring z values through equation (4), as such affecting the 17 survival and transition probabilities S and P. However, the recapture probabilities R are 18 calculated directly from the data and are therefore not affected by z. 19 Our prior assumptions followed the robust design framework (Pollock 1982): 20 since secondary capture sessions were very close to each other (1 day), we assumed that 21 survival probability was one between the secondary sessions and no transitions were 22 possible between them. These priors were implicit in our computations, since we 3 1 estimated survival and transition only between primary capture sessions. For all survival 2 and transition estimates among years, we used uniform priors, spanning all possible 3 probabilities (0 to 1 interval). Thus, from our ten-year capture histories, we estimated 4 posterior probability distributions for annual survival, recapture, and transition 5 probabilities. 6 The total likelihood for observing data x and having true life stages z, given the 7 vital rate parameters, was obtained by taking a product of equations (1-3) over years. The 8 posterior distributions for the vital rates can be written down analytically (beta 9 distribution for recapture and survival and Dirichlet distribution for transition). Thus, 10 given our observed data x and an estimate for life stages z, we sampled new values for R, 11 S, and P from these known distributions. In contrast, for some given values of the vital 12 rates, we calculated the probabilities for different zit using equation (4) and sampled new 13 values from the resulting multinomial distribution. Alternating these conditional 14 samplings, we arrived at the Gibbs sampler (Gelman et al. 2004), which can be 15 summarized by the following four steps: 16 1. Set the initial values for R, S, P and z and iteration counter i=1 17 2. Sample R, S and P from their conditional posteriors, given x and z, using 18 19 20 21 equations (1-3) 3. Loop over individuals and time points where xkt = 0, sample a new life stage zkt using equation (4), given previously sampled R, S, and P 4. Set i=i+1 and go to step 2, until a desired number of iterations is obtained 22 We determined that the Gibbs sampler had converged by checking that the chains had 23 mixed well. Additionally, we observed that the likelihood stabilized after an initial period 4 1 of increase, which indicated convergence. We produced 5000 samples from the posterior. 2 In addition to the above survival and stage transitions, we estimated two 3 additional terms, breeding probability and survival probability from egg to one year. For 4 the breeding probability, we estimated of the number of females for each year. These 5 were calculated by dividing the number of observed females (counted from the field data) 6 by the recapture probability estimate. To get the estimate for breeding probability as a 7 distribution, this calculation was performed by numerical integration over the distribution 8 of recapture rates given by the Bayesian mark-recapture analysis, i.e., repeating 9 calculations over the distribution of estimates generated by the Gibbs sampler. We then 10 divided the number of observed egg masses by the number of females to determine the 11 proportion of females breeding each year. 12 Similarly, for survival probabilities from egg to one year, we needed estimates of 13 the number of one-year-old frogs. These were calculated from a separate data set, which 14 contains capture histories for one-year-old individuals (identified as frogs 34 mm and 15 smaller). The above Gibbs sampling algorithm was used for estimating the recapture 16 probability for this age class (in this estimation, we only had one life stage), and the 17 number of observed individuals was then divided by different recapture probability 18 estimates to get a distribution of the number of one-year-old frogs. To estimate survival 19 from egg to one year, we then divided the number of one-year-old frogs by the number of 20 eggs laid in the basin the previous spring. 21 22 23 24 5 1 Literature Cited 2 3 Calvert, A. M., Bonner, S. J., Jonsen, I. D., Flemming, J. M., Walde, S. J. & Taylor, P. D. 4 (2009) A hierarchical Bayesian approach to multi-state mark-recapture: 5 simulations and applications. Journal of Applied Ecology, 46, 610-620. 6 Clark, J. S., Ferraz, G., Oguge, N., Hays, H. & DiConstanzo, J. (2005) Hierarchical 7 Bayes for structured, variable populations: from recapture data to life-history 8 prediction. Ecology, 86, 2232-2244. 9 10 11 12 13 14 Dupuis, J. A. (1995) Bayesian estimation of movement and survival probabilities from capture-recapture data. Biometrika, 82, 761-772. Dupuis, J. A. & Schwarz, C. J. (2007) A Bayesian approach to the multistate Jolly-Seber capture-recapture model. Biometrics, 63, 1015-1022. Gelman, A., Carlin, J. B., Stern, H. S. & Rubin, D. B. (2004) Bayesian data analysis (2nd ed.), Chapman and Hall/CRC, Boca Raton. 15 McCaffery, R. M. & Maxell, B. A. (2010) Decreased winter severity increases viability 16 of a montane frog population. Proceedings of the National Academy of Sciences 17 107, 8644-8649. 18 19 20 21 22 Royle, J. A. (2008) Modeling individual effects in the Cormack-Jolly-Seber model: a state-space formulation. Biometrics, 64, 364-370. Royle, J. A. & Dorazio, R. M. (2008) Hierarchical modeling and inference in ecology, Academic Press, San Diego. Royle, J. A. & Dorazio, R. M. (2010) Parameter-expanded data augmentation for 23 Bayesian analysis of capture-recapture models. Journal of Ornithology, DOI: 24 10.1007/s10336-010-0619-4. 6 1 2 3 Royle, J. A. & Kery, M. (2007) A Bayesian state-space formulation of dynamic occupancy models. Ecology, 88, 1813-1823. Zheng, C., Ovaskainen, O., Saastamoinen, M. & Hanski, I. (2007) Age-dependent 4 survival analyzed with Baysian models of mark-recapture data. Ecology, 88, 5 1970-1976. 6 7 7