GSA 2017 NA mammal species pool

My GSA 2017 talk on the multi-level dynamics of species pool functional diversity.

Transcript

1. Modeling changes to the functional composition of North American mammal diveristy multi-level dynamics of a regional species pool Peter D Smits Department of Integrative Biology, University of California – Berkeley
2. Question Why do the relative diversities of functional groups change within a species pool? function of species traits and environmental context
3. Eco-cube and functional groups (Bambach et al., 2007, Palaeontology)
4. Species pool concept (Mittelbach and Schemske, 2015, TREE)
5. Structured, multi-level data in biology (Ovaskainen et al. 2017 Ecology Letters)
6. Cenozoic mammals of North America (Rudolph Zallinger)
7. Conceptualizing the knowns and unknowns traits trait species true presence time species environment time factor observed presence species time “observed” data important missing information true relative diversity effects + correlations between traits/factors
8. Covariates of interest, temporal structure species occurrence (∼1400 species) per NALMA functional group dietary category: carnivore, herbivore, insectivore, omnivore locomotor category: arboreal, digitigrade, fossorial, plantigrade, scansorial, unguligrade observation indiv-level: species functional group mean mass time of observation origination indiv-level: species functional group taxon order mean mass group-level: FG/time temperature est Mg/Ca plant phase (Pa-Eo, Eo-Mi, Mi-Pl) survival indiv-level: species functional group taxon order mean mass group-level: FG/time temperature est Mg/Ca plant phase (Pa-Eo, Eo-Mi, Mi-Pl)
9. Conceptualizing the analysis true presence observed presence observation probability traits environment trait species species time time time factor species time trait species select traits
10. Hidden Markov Model with absorbing state Jolly-Seber CMR/Restricted occupancy model yi,t ∼ Bernoulli(zi,tpi,t) zi,t=1 ∼ Bernoulli(φi,t=1) zi,t ∼ Bernoulli zi,t−1πi,t + t x=1 (1 − zi,x )φi,t y observed state; z estimated state p observation; φ origination; π survival i in N; t in T
11. Modeling the probabilities; individual-level Multi-level logistic regression pi,t ∼ logit−1(bt + ej[i] + βpmassi ) φi,t ∼ logit−1(f φ j[i],t + oφ k[i] + βφmassi ) πi,t ∼ logit−1(f π j[i],t + oπ k[i] + βπmassi ) observation: bt time-varying intercept; ej[i] functional group eff; βp mass eff origination: f φ j[i],t time/FG-varying intercept; oφ j[i] order eff; βφ mass eff survival: f π j[i],t time/FG-varying intercept; oπ j[i] order eff; βπ mass eff
12. Modeling the probabilities; group-level Multivariate regression of time/FG-varying intercept f φ ∼ MVN    Uγφ j=1 . . . Uγφ j=J , diag(τf φ )Ωf φ diag(τf φ )    f π ∼ MVN    Uγπ j=1 . . . Uγπ j=J , diag(τf π )Ωf π diag(τf π )    U matrix group-level covariates; γφ, γπ vectors group-level reg coefs Ωφ, Ωπ corr matrix of FG by time; τφ, τπ scale of FG by time
13. Modeling the probabilities; final details Comments on priors, implementation random-walk priors on time-varying intercepts regularizing priors with some specific predictions very weak/no effect of mass e.g. N(0, 0.5) very weak/no effect of group-level covariates e.g. N(0, 0.5) very weak/no correlation b/w functional groups e.g. LKJ(2) marginalization problem b/c gradient based estimation
14. Parameter estimation and inference Bayesian inference intuitive and expressive regularization/partial pooling external information Automatic Differentiation Variational Inference (ADVI) when full HMC/MCMC slow approx Bayesian inference; assumes posterior is Gaussian true Bayesian posterior Stan
15. Model adequate? Posterior predictive check mean # occurrences per species from datasets simulated from posterior
16. Observation; NALMA log-odds of observing a species, given that it is present
17. Observation; functional group log-odds of observing a species, given that it is present
18. Origination; individual-level probability of species originating, given it hasn’t originated yet
19. Origination; group-level change to log-odds of species originating, given it hasn’t originated yet
20. Survival; individual-level probability of species surviving, given it was present
21. Survival; group-level change to log-odds of species surviving, given it was present
22. Standing diversity of functional groups through time
23. Relative diversity of functional groups through time
24. Changes to relative diversity between Neogene/Paleogene increase digitigrade, plantigrade, unguligrade herbivores fossorial functional groups plantigrade omnivores decrease near total loss of arboreal functional groups plantigrade, scansorial insectivores unguligrade omnivores
25. Conclusions temporal differences have larger effect on P(observation) than effect of FG increase in P(origination) often met with decrease in P(survival), but not 1-to-1 environmental covariates effect origination within FG more often than survival no evidence for correlation in origination or survival of functional groups over time that is not accounted for by RW prior potential for short-term similarity, just no long-term correlation HMC/MCMC might tweak these results b/c ADVI assumptions (Gaussian posterior)
26. Acknowledgements UC Berkeley Seth Finnegan, Adiel Klompmaker, Emily Orzechowski, Larry Taylor, Sara Kahanamoku, Josh Zimmt UChicago Kenneth D. Angielczyk, Michael J. Foote, P. David Polly, Richard H. Ree, Graham Slater psmits.github.io/ coping @PeterDSmits