README.md

In sustainablefarms/msod: Multispecies Occupancy Detection Models

msod (Multi-Species Occupancy Detection Modelling)

This package is in draft form: it is in the early stage of being designed

Package for Bayesian fitting of a small number of occupancy detection models for multiple species. Models are closely related to a JSDM by Tobler and models seen in the Boral package.

Code Structure

• Computations Common to All Models
• linear combinations of environmental occupancy predictors for many parameter draws
• linear combination of LV
• apply a modelsite x theta --> likelihood function to get array of likelihood of sites and theta
  • probably modelsite x theta x species arrays, with each value a probability of occupancy (detection), conditional on certain parameters, marginal on others.
• combine occupancy probability with conditional detection probability to get full detection likilihoods (full probability distribution has too many dimensions) [doing this in multisite way - e.g. through large arrays]
• computing a median, hpd, and expectation for occupancy from given occupancy probability per theta x LV x RE
• computing a median, hpd, and expectation for expected species richness given occupancy probability per theta x LV x RE
• processing of environmental data, and their saving to a fitted model

• common arrays:

  • value: prob occupancy conditional on some, marginal on others dim: draw x modelsite x species x marginal param draw
  • value: prob detection conditional on occupancy, marginal on others. dim: draw x modelsite x species x marginal param draw
  • value: stdnormthresh for occupancy with marginal param sets. dim: draw x modelsite x species x marginal param draw

• Computations that will differ between models

• poccupy() and poccupy_raw(): occupancy RV mean and standard deviation (it will always be normal though)
  • mean will be sum of external predictor component, LV component and RE component depending on the model
  • functions that specific to each model, but have similar interface:
  • arguments are covariate values, random covariate values (to be marginalised across perhaps), and their loadings. Consider making loading all the one array, with a second argument indicating which of the loadings are the random covariate values.
  • fixed covariate values are fixed for each model site (even if they are estimated) [this is a modelsite x covariate matrix]. Corresponding loadings are drawn from the posterior, any number of draws are allowed [this is a species x covariate x draw array].
  • random covariate values there is a sample of these for each model site [this is a modelsite x covariate x draw array], loadings correspond to each of the fixed covariate loading draws [this is a species x covariate x draw array].
  • additional options specific to each model (e.g. including LV uncertainty using theory rather than simulation)
  • RETURNS: array of occupancy probability values modelsite x species x draw
  • have a default version that is applied to a fitted object
• pdet_occ() and pdet_occ_raw(): detection prob when occupied (but only different for some models)
  • same interface style to the occupancy predictions, except returns visit x species x draw
• likelihood(): likelihood similar to above probabilities but also needs observed values
• functions for marginalising modelsite x species x draw
• prepJAGS2(): parameters to check and initialise for running JAGS

• preparing random effects

• Model Classes

• inheritance structure: runjags > JSODM without LV > JSODM with LV > (JSODM with LV and RE, JSODM with spatially correlated LV). Actually this seems poor, I think starting with broadest and going more narrow (narrowest == JSODM witout LV) would be more appropriate.
• Use S3 classes (I found this page useful https://elvinouyang.github.io/study%20notes/oop-with-r-s3-and-r6/)
  • generic functions:
  • logLik
  • (predict) compute probabilities
  • simulate
  • prepdata (missing values, LV and random effects parts)

Functionality (draft)

Core

• fitting JSDMs with occupancy and detection components
• latent variables
• spatial correlated latent variables
• random effects
• probit tranfer function with linear predictor for occupancy
• logit model for detection

• some missing data

• computing likelihood and log posterior density (lpd) for each ModelSite

• building models with artifical parameters (core because very important for code testing)

• predicting for multiple species and multiple sites

• simulating from given models (core because very important for code testing)

• remove the standardising of data to 1 with mean 0

Small Helping Functionality

• standardising and preparing data
• ViF removal of highly correlated predictors
• Rmd template for assessing model quality

Non-Core Functionality (may not be available for all models)

• Occupancy and Detection Residuals
• Species richness predictions and comparison to observed

sustainablefarms/msod documentation built on March 6, 2023, 7:17 a.m.