Happy to finally see these results out! Project started 4 yrs ago from discussions with @noellebeckman at @sesync, study made possible with data from @stri_panama and resources from @UQAT & @ComputeCanada.
I would like to mention how technical challenges could be solved with the tools made available by the @mcmc_stan team, and @betanalpha's case studies illustrating best practices for Bayesian analysis: betanalpha.github.io/assets/case_st… . That story is mostly in Appendix 2 of our paper.
The main challenge in estimating #seeddispersal kernels is that they are often very leptokurtic, i.e. many seeds falling either very close or quite far from the parent plant. Seed sampling data has both an upper bound in terms of observed distances from parents...
but also a lower bound, especially in a highly diverse forest; i.e. with a limited number of seed traps, they cannot be very close to parents of every species. As a result, the data alone cannot constrain the parameters of many potential kernel functions.
Even #Bayesian approaches to the problem have often chosen specific kernel functions or set hard bounds on parameters values (bounded uniform distribution) based on identifiability constraints. In this study, we considered a wider range of possible dispersal kernels...
and chose weakly-uniform priors based on summary statistics of the resulting kernels, e.g. it is very unlikely for mean dispersal distances to be <1m and >1km. We also took advantage of the recent "loo" package of @avehtari et al. to produce model-averaged dispersal curves.
Briefly: The overall strategy is to embrace the uncertainty about the functional form of the kernel and to include it to predict the biologically significant quantities, i.e. how many individuals disperse at different distances. The Bayesian approach is very well-suited to this.
