Fit vector autoregressive spatio-temporal model

Fits a vector autoregressive spatio-temporal (VAST) model using a minimal feature-set and a widely used interface.

Usage

tinyVAST(
  formula,
  data,
  time_term = NULL,
  space_term = NULL,
  spacetime_term = NULL,
  family = gaussian(),
  space_columns = c("x", "y"),
  spatial_domain = NULL,
  time_column = "time",
  times = NULL,
  variable_column = "var",
  variables = NULL,
  distribution_column = "dist",
  delta_options = list(formula = ~1),
  spatial_varying = NULL,
  weights = NULL,
  control = tinyVASTcontrol(),
  ...
)

Arguments

formula: Formula with response on left-hand-side and predictors on right-hand-side, parsed by mgcv and hence allowing s(.) for splines or offset(.) for an offset.
data: Data-frame of predictor, response, and offset variables. Also includes variables that specify space, time, variables, and the distribution for samples, as identified by arguments variable_column, time_column, space_columns, and distribution_column.
time_term: Specification for time-series structural equation model structure for constructing a time-variable interaction that defines a time-varying intercept for each variable (i.e., applies uniformly across space). time_term=NULL disables the space-variable interaction; see make_dsem_ram() for notation.
space_term: Specification for structural equation model structure for constructing a space-variable interaction. space_term=NULL disables the space-variable interaction; see make_sem_ram() for notation.
spacetime_term: Specification for time-series structural equation model structure including lagged or simultaneous effects for constructing a time-variable interaction, which is then combined in a separable process with the spatial correlation to form a space-time-variable interaction (i.e., the interaction occurs locally at each site). spacetime_term=NULL disables the space-variable interaction; see make_dsem_ram() or make_eof_ram().
family: A function returning a class family, including gaussian(), lognormal(), tweedie(), binomial(), Gamma(), poisson(), nbinom1(), or nbinom2(). Alternatively, can be a named list of these functions, with names that match levels of data$distribution_column to allow different families by row of data. Delta model families are possible, and see Families for delta-model options,
space_columns: A string or character vector that indicates the column(s) of data indicating the location of each sample. When spatial_domain is an igraph object, space_columns is a string with with levels matching the names of vertices of that object. When spatial_domain is an fmesher or sfnetwork object, space_columns is a character vector indicating columns of data with coordinates for each sample.
spatial_domain: Object that represents spatial relationships, either using fmesher::fm_mesh_2d() to apply the SPDE method, igraph::make_empty_graph() for independent time-series, igraph::make_graph() to apply a simultaneous autoregressive (SAR) process, sfnetwork_mesh() for stream networks, or NULL to specify a single site. If using igraph then the graph must have vertex names V(graph)$name that match levels of data[,'space_columns']
time_column: A character string indicating the column of data listing the time-interval for each sample, from the set of times in argument times.
times: A integer vector listing the set of times in order. If times=NULL, then it is filled in as the vector of integers from the minimum to maximum value of data$time. Alternatively, it could be the minimum value of data$time through future years, such that the model can forecast those future years.
variable_column: A character string indicating the column of data listing the variable for each sample, from the set of times in argument variables.
variables: A character vector listing the set of variables. if variables=NULL, then it is filled in as the unique values from data$variable_columns.
distribution_column: A character string indicating the column of data listing the distribution for each sample, from the set of names in argument family. if variables=NULL, then it is filled in as the unique values from data$variables.
delta_options: a named list with slots for formula, space_term, and spacetime_term. These specify options for the second linear predictor of a delta model, and are only used (or estimable) when a delta family is used for some samples.
spatial_varying: a formula specifying spatially varying coefficients.
weights: A numeric vector representing optional likelihood weights for the data likelihood. Weights do not have to sum to one and are not internally modified. Thee weights argument needs to be a vector and not a name of the variable in the data frame.
control: Output from tinyVASTcontrol(), used to define user settings.
...: Not used.

Value

An object (list) of class tinyVAST. Elements include:

data: Data-frame supplied during model fitting
spatial_domain: the spatial domain supplied during fitting
formula: the formula specified during model fitting
obj: The TMB object from MakeADFun
opt: The output from nlminb
opt: The report from obj$report()
sdrep: The output from sdreport
tmb_inputs: The list of inputs passed to MakeADFun
call: A record of the function call
run_time: Total time to run model
interal: Objects useful for package function, i.e., all arguments passed during the call
deviance_explained: output from deviance_explained

Details

tinyVAST includes four basic inputs that specify the model structure:

formula specifies covariates and splines in a Generalized Additive Model;
space_term specifies interactions among variables and over time, constructing the space-variable interaction.
spacetime_term specifies interactions among variables and over time, constructing the space-time-variable interaction.
spatial_domain specifies spatial correlations

the default spacetime_term=NULL and space_term=NULL turns off all multivariate and temporal indexing, such that spatial_domain is then ignored, and the model collapses to a generalized additive model using gam. To specify a univariate spatial model, the user must specify spatial_domain and either space_term="" or spacetime_term="", where the latter two are then parsed to include a single exogenous variance for the single variable

Model type	How to specify
Generalized additive model	specify `spatial_domain=NULL` `space_term=""` and `spacetime_term=""`, and then use `formula` to specify splines and covariates
Dynamic structural equation model (including vector autoregressive, dynamic factor analysis, ARIMA, and structural equation models)	specify `spatial_domain=NULL` and use `spacetime_term` to specify interactions among variables and over time
Univariate spatio-temporal model, or multiple independence spatio-temporal variables	specify `spatial_domain` and `spacetime_term=""`, where the latter is then parsed to include a single exogenous variance for the single variable
Multivariate spatial model including interactions	specify `spatial_domain` and use `space_term` to specify spatial interactions
Vector autoregressive spatio-temporal model (i.e., lag-1 interactions among variables)	specify `spatial_domain` and use `spacetime_term=""` to specify interactions among variables and over time, where spatio-temporal variables are constructed via the separable interaction of `spacetime_term` and `spatial_domain`

Examples

# Simulate a seperable two-dimensional AR1 spatial process
n_x = n_y = 25
n_w = 10
R_xx = exp(-0.4 * abs(outer(1:n_x, 1:n_x, FUN="-")) )
R_yy = exp(-0.4 * abs(outer(1:n_y, 1:n_y, FUN="-")) )
z = mvtnorm::rmvnorm(1, sigma=kronecker(R_xx,R_yy) )

# Simulate nuissance parameter z from oscillatory (day-night) process
w = sample(1:n_w, replace=TRUE, size=length(z))
Data = data.frame( expand.grid(x=1:n_x, y=1:n_y), w=w, z=as.vector(z) + cos(w/n_w*2*pi))
Data$n = Data$z + rnorm(nrow(Data), sd=1)

# Add columns for multivariate and/or temporal dimensions
Data$var = "n"

# make SPDE mesh for spatial term
mesh = fmesher::fm_mesh_2d( Data[,c('x','y')], n=100 )

# fit model with cyclic confounder as GAM term
out = tinyVAST( data = Data,
                formula = n ~ s(w),
                spatial_domain = mesh,
                space_term = "n <-> n, sd_n" )

# Run crossvalidation
CV = cv::cv( out )
#> R RNG seed set to 266497
#> Error in eval(call, parent.frame()): object 'mesh' not found
CV
#> Error: object 'CV' not found