search

brianstock / MixSIAR

A framework for Bayesian mixing models in R:
http://brianstock.github.io/MixSIAR/
91 stars 75 forks source link

Request for guidance regarding the R-MixSIAR package #174

Closed brianstock closed 5 years ago

brianstock commented 5 years ago

I’m very interested in using your R-package MixSIAR in my research that focuses on vegetation source water identification. I have a few questions and I’m just wondering if I may request your guidance in addressing them.

  1. How to generate discrimination data from a given problem? For example, would you please let me know if I may request a step-by-step procedure that you followed to get the values listed in the file geese_discrimination.csv (example 3 in the user’s manual)?

  2. Let us suppose we have mean and one standard deviation for all the sources in a mixing model. Would you please let me know what method is used to sample the parameter space for the sources for creating the probability density function (PDF) for the mixture? Also, would you please let me know if there are any inherent assumptions involved regarding the parameter distribution for any sources for creating the pdf for source mixtures?

Thank you!

brianstock commented 5 years ago

  1. I'm not an expert on this, especially the plant water sourcing application - look in the methods section of papers that use mixing models on problems you're interested in (plant water sourcing). The geese example data are from Inger et al. 2006. They did not know the true values and conducted a sensitivity analysis: "A sensitivity analysis was carried out on the fractionation factors. We varied the value used for δ15N between 3 and 5, and the value used for δ13C between 1 and 2." Given that mixing model results can be sensitive to fractionation values, it can make sense to do this even if you have a good idea what the values should be. For animal diet applications, a group has made an add-on model, SIDER, that takes phylogeny into account. Reading that paper and references cited within will give you a sense of what people do for animal diet applications. Best are feeding experiments. An example: for mixing model analysis in deVries et al. 2016, they used experimentally-derived fractionation described in deVries et al. 2015).

  2. MixSIAR uses MCMC to sample the parameter space. Main advantage of this approach over previous, ad-hoc methods (e.g. IsoSource), is that uncertainty is included in a probabilistic model. The first Bayesian mixing model papers spelled this out (Moore & Semmens 2008, Parnell et al. 2010). The MixSIAR model and parameterization is described in Stock et al. 2018 (also see supplement and user manual).

Hope that helps!