latent_space: Latent social space models for interaction and association...

In MNWeiss/aninet: Statistical Models for Animal Social Networks

Latent social space models for interaction and association data

Description

Fits a D dimensional latent space model of the given family to a matrix of associations or interactions

Usage

latent_space(
  formula,
  family = "poisson",
  dimensions = 2,
  ind.RE = T,
  effort,
  beta.prior = c(0, 1e-04),
  vcv.prior = NULL,
  re.prior = c(0.1, 0.1),
  z0 = NULL,
  ...
)

Arguments

formula	A glm style formula, where the response and predictors are all square matrices of the same dimension. The response matrix should be the counts of interactions or associations.
family	Character, one of either "poisson" or "binomial". See details.
dimensions	Integer, the number of latent dimensions to model.
ind.RE	Logigal, indicating whether to include an individual sociality random effect.
effort	A square matrix indicating dyadic sampling effort. See details.
beta.prior	Numeric vector, prior for fixed effects. Should be the mean and precision of a normal distribution.
vcv.prior	Numeric matrix, prior for the variance-covariance matrix of latent positions. Should be a matrix with row and column number equal to the dimensions, representing a parameterization of the Wishart distribution.
re.prior	Numeric vector, prior for individual random effect. Should be the parameters of a Gamma distribution.
z0	Optional numeric matrix, initial positions of nodes in the latent space. If not provided, initial values are generated using mutlidimensional scaling.
...	Further arguments to be passed to runjags.

Details

Social networks often exhibit transitivity, where a connection between B and C is likely to be stronger if B and C have strong connections to a third individual A. This tendency is often referred to as "triadic closure." One way to account for this in regression settings is to view nodes as being placed within a D dimensional latent space, with edges partially determined by the euclidean distances between nodes. This function fits one of these latent space models to a matrix of associations or interactions. For interactions, a Poisson model should be fit, with effort indicating the sampling time per dyad. For associations, a binomial model should be fit, with effort being indicated by the denominator of the association index. In both cases, the response matrix should be a matrix of integers, indicating the number of dyadic interactions or associations. This function fits the model using Gibbs sampling via JAGS and runjags, and allows the user to define priors for fixed effects, the covariance matrix of the latent positions, and individual random effects (if included).

Value

A named list containing the following slots

Slots

summary

The summary statistics for the fixed effects and variance of the random effects (if included).

distances

The sampled distances between nodes in social space, stored as a S x N x N array, where S is the number of MCMC samples.

z

The sampled positions of nodes in social space, stored as a S x N x D array, where S is the number of MCMC samples and D is the number of latent dimensions.

jags_model

The full runjags object, which can be used to assess convergence through trace plots and diagnostics, and to get the DIC value of the model.

MNWeiss/aninet documentation built on Jan. 31, 2023, 3:55 a.m.