mbm: Create an mbm model
In mtalluto/mbmtools: Multifaceted Biodiversity Modelling

Description Usage Arguments Details Value References

Create an mbm model

mbm(y, x, y_name = "beta", link = c("identity", "probit"),
  likelihood = c("gaussian"), lengthscale = NULL, sparse = FALSE,
  force_increasing = FALSE, sparse_inducing = 10, sparse_batch = 10,
  sparse_iter = 10000, exact_thresh = 100, verbose = FALSE)

`y`	Square dissimilarity or distance matrix, can be complete or lower triangular only. Row and column names are required and must match the site names in the rows of `x`.
`x`	Matrix giving a series of covariates (in columns) for all sites (in rows). Row names are required. All variables will be included in the model.
`y_name`	A name to give to the y variable
`link`	Link function to use
`likelihood`	Likelihood function to use
`lengthscale`	Either NULL (in which case all lengthscales will be optimized) or a numeric vector of length `ncol(x)+1`. If a vector, the first entry corresponds to environmental distance, and entries `i = 1 + (1:n)` correspond to the variable in x[,i]. Values must be `NULL` or positive numbers; if NULL, the corresponding lengthscale will be set via optimization, otherwise it will be fixed to the value given.
`sparse`	Should we use the stochastic variational GP (see 'details').
`force_increasing`	Boolean; if true, beta diversity will be constrained to increase with environmental distance
`sparse_inducing`	Number of inducing inputs to use if 'sparse = TRUE'
`sparse_batch`	Batch size to use if 'sparse = TRUE'
`sparse_iter`	Maximum number of optimizer iterations if 'sparse = TRUE'
`exact_thresh`	integer; threshold at which mbm will refuse to run an exact gp.
`verbose`	Should messages during model fitting be printed?

For larger datasets (more than ~100 sites), it is recommended to use 'sparse=TRUE'. This will use a sparse approximation to the default method, following the stochastical variational GP (Hensman et al 2013). Note that if a link function is selected, it will be applied as a transformation of the y data–i.e., for link function L() we fit a SVGP to describe the expectation E(L(y))–rather than as a true link function– fitting L(E(y))–as is done when 'svgp=FALSE'. This is due to a limitation in the underlying GP library.

An S3 object of class mbm, containing the following components: * 'x': the original (untransformed) site by covariate matrix * 'y': the original (untransformed) site by site diversity data * 'covariates': Transformed x-variables supplied to mbm * 'response': Transformed response variable; this is the data supplied to mbm * 'covar_sites': Site names to match the covariate matrix * 'y_transform': transformation applied to y-data before modelling * 'y_rev_transform': reverse transformation to get y-data back on the original scale * 'link': a character string identifying the link function * 'inv_link': inverse of the link function * 'pyobj': A list of python objects used by the model; this is not meant for user interaction

Hensman J, Fusi N, and Lawrence ND. 2013. Gaussian Processes for Big Data. In: In Proceedings of the 29th Conference on Uncertainty in Artificial Intelligence.

mtalluto/mbmtools documentation built on Aug. 13, 2019, 9:44 a.m.