bru_obs: Observation model construction for usage with 'bru()'

In fbachl/inlabru: Bayesian Latent Gaussian Modelling using INLA and Extensions

Observation model construction for usage with bru()

Description

Observation model construction for usage with bru().

Note: Prior to version ⁠2.12.0⁠, this function was called like(), and that alias will remain for a while until examples etc have been updated and users made aware of the change. The name change is to avoid issues with namespace clashes, e.g. with data.table::like(), and also to signal that the function defines observation models, not just likelihood functions.

Usage

bru_obs(
  formula = . ~ .,
  family = "gaussian",
  data = NULL,
  response_data = NULL,
  data_extra = NULL,
  E = NULL,
  Ntrials = NULL,
  weights = NULL,
  scale = NULL,
  domain = NULL,
  samplers = NULL,
  ips = NULL,
  used = NULL,
  allow_combine = NULL,
  aggregate = NULL,
  aggregate_input = NULL,
  control.family = NULL,
  tag = NULL,
  options = list(),
  .envir = parent.frame(),
  include = deprecated(),
  exclude = deprecated(),
  include_latent = deprecated()
)

like(
  formula = . ~ .,
  family = "gaussian",
  data = NULL,
  response_data = NULL,
  E = NULL,
  Ntrials = NULL,
  weights = NULL,
  scale = NULL,
  domain = NULL,
  samplers = NULL,
  ips = NULL,
  used = NULL,
  allow_combine = NULL,
  control.family = NULL,
  tag = NULL,
  options = list(),
  .envir = parent.frame(),
  mesh = deprecated(),
  include = deprecated(),
  exclude = deprecated(),
  include_latent = deprecated()
)

bru_obs_list(...)

## S3 method for class 'list'
bru_obs_list(object, ..., .envir = NULL)

## S3 method for class 'bru_obs_list'
bru_obs_list(..., .envir = NULL)

## S3 method for class 'bru_obs'
c(..., .envir = NULL)

## S3 method for class 'bru_obs_list'
c(..., .envir = NULL)

## S3 method for class 'bru_obs_list'
x[i]

Arguments

formula	a formula where the right hand side is a general R expression defines the predictor used in the model.
family	A string identifying a valid INLA::inla likelihood family. The default is gaussian with identity link. In addition to the likelihoods provided by inla (see names(INLA::inla.models()$likelihood)) inlabru supports fitting latent Gaussian Cox processes via family = "cp". As an alternative to bru(), the lgcp() function provides a convenient interface to fitting Cox processes.
data	Predictor expression-specific data, as a data.frame, sf, or SpatialPoints[DataFrame] object.
response_data	Observation/response-specific data for models that need different size/format for inputs and response variables, as a data.frame, sf, or SpatialPoints[DataFrame] object.
data_extra	object convertible with as.list() with additional variables to be made available in predictor evaluations. Variables with the same names as the data object will be ignored, unless accessed via .data_extra.[["name"]] or .data_extra.$name in the formula.
E	Exposure/effort parameter for family = 'poisson' passed on to INLA::inla. Special case if family is 'cp': rescale all integration weights by a scalar E. For sampler specific reweighting/effort, use a weight column in the samplers object instead, see fmesher::fm_int(). Default taken from options$E, normally 1.
Ntrials	A vector containing the number of trials for the 'binomial' likelihood. Default taken from options$Ntrials, normally 1.
weights	Fixed (optional) weights parameters of the likelihood, so the log-likelihood⁠[i]⁠ is changed into weights[i] * log_likelihood[i]. Default value is 1. WARNING: The normalizing constant for the likelihood is NOT recomputed, so ALL marginals (and the marginal likelihood) must be interpreted with great care. For family = "cp", the weights are applied as sum(weights * eta) in the point location contribution part of the log-likelihood, where eta is the linear predictor, and do not affect the integration part of the likelihood. This can be used to implement approximative methods for point location uncertainty.
scale	Fixed (optional) scale parameters of the precision for several models, such as Gaussian and student-t response models.
domain, samplers, ips	Arguments used for family="cp". domain Named list of domain definitions. samplers Integration subdomain for 'cp' family. ips Integration points for 'cp' family. Defaults to fmesher::fm_int(domain, samplers). If explicitly given, overrides domain and samplers.
used	Either NULL (default) or a bru_used() object, that overrides the include, exclude, include_latent arguments. When used is NULL (default), the information about what effects and latent vectors are made available to the predictor evaluation is defined by used <- bru_used( formula, effect = include, effect_exclude = exclude, latent = include_latent )
allow_combine	logical; If TRUE, the predictor expression may involve several rows of the input data to influence the same row. When NULL, defaults to FALSE, unless response_data is non-NULL, or data is a list, or the likelihood construction requires it.
aggregate	character ("none", "sum", "average", "logsumexp", or "logaverageexp", as defined by bru_mapper_aggregate(type = aggregate)) or an aggregation bru_mapper object (bru_mapper_aggregate() or bru_mapper_logsumexp()). Default NULL, interpreted as "none". , available from version ⁠2.12.0.9013⁠.
aggregate_input	NULL or an optional input list to the mapper defined by non-NULL aggregate, overriding the default, list(block = .data.[[".block"]], weights = .data.[["weight"]], n_block = bru_response_size(.response_data.)) , available from version ⁠2.12.0.9013⁠.
control.family	A optional list of INLA::control.family options
tag	character; Name that can be used to identify the relevant parts of INLA predictor vector output, via bru_index().
options	A bru_options options object or a list of options passed on to bru_options()
.envir	The evaluation environment to use for special arguments (E, Ntrials, weights, and scale) if not found in response_data or data. Defaults to the calling environment.
include, exclude, include_latent	, use used instead. Arguments controlling what components and effects are available for use in the predictor expression. include Character vector of component labels that are used as effects by the predictor expression; If NULL (default), the bru_used() method is used to extract the variable names from the formula. exclude Character vector of component labels to be excluded from the effect list determined by the include argument. Default is NULL; do not remove any components from the inclusion list. include_latent Character vector. Specifies which latent state variables are directly available to the predictor expression, with a ⁠_latent⁠ suffix. This also makes evaluator functions with suffix ⁠_eval⁠ available, taking parameters main, group, and replicate, taking values for where to evaluate the component effect that are different than those defined in the component definition itself (see bru_comp_eval()). If NULL, the bru_used() method auto-detects use of ⁠_latent⁠ and ⁠_eval⁠ in the predictor expression.
mesh	Ignored.
...	For bru_obs_list.bru_obs, one or more bru_obs objects
object	A list of bru_obs objects
x	bru_obs_list object from which to extract element(s)
i	indices specifying elements to extract

Value

A likelihood configuration which can be used to parameterise bru().

Methods (by generic)

• c(bru_obs): Combine several bru_obs objects into a bru_obs_list object

Functions

• like(): Legacy like() method for inlabru prior to version ⁠2.12.0⁠. Use bru_obs() instead.

• bru_obs_list(): Combine bru_obs observation model object into a bru_obs_list object

• bru_obs_list(list): Combine one or more lists of bru_obs observation model objects into a bru_obs_list object

• bru_obs_list(bru_obs_list): Combine a list of bru_obs observation model objects into a bru_obs_list object

• c(bru_obs_list): Combine several bru_obs_list objects into a bru_obs_list object

Author(s)

Fabian E. Bachl [e-mail]

Finn Lindgren [e-mail]

See Also

bru_response_size(), bru_used(), bru_comp(), bru_comp_eval()

summary.bru_obs()

Examples

if (bru_safe_inla() &&
    require(ggplot2, quietly = TRUE)) {

  # The 'bru_obs()' (previously 'like()') function's main purpose is to set up
  # observation models, both for single- and multi-likelihood models.
  # The following example generates some random covariates which are observed
  # through two different random effect models with different likelihoods

  # Generate the data

  set.seed(123)

  n1 <- 200
  n2 <- 10

  x1 <- runif(n1)
  x2 <- runif(n2)
  z2 <- runif(n2)

  y1 <- rnorm(n1, mean = 2 * x1 + 3)
  y2 <- rpois(n2, lambda = exp(2 * x2 + z2 + 3))

  df1 <- data.frame(y = y1, x = x1)
  df2 <- data.frame(y = y2, x = x2, z = z2)

  # Single likelihood models and inference using bru are done via

  cmp1 <- y ~ -1 + Intercept(1) + x
  fit1 <- bru(cmp1, family = "gaussian", data = df1)
  summary(fit1)

  cmp2 <- y ~ -1 + Intercept(1) + x + z
  fit2 <- bru(cmp2, family = "poisson", data = df2)
  summary(fit2)

  # A joint model has two likelihoods, which are set up using the bru_obs
  # function

  lik1 <- bru_obs(
    "gaussian",
    formula = y ~ x + Intercept,
    data = df1,
    tag = "norm"
  )
  lik2 <- bru_obs(
    "poisson",
    formula = y ~ x + z + Intercept,
    data = df2,
    tag = "pois"
  )

  # The union of effects of both models gives the components needed to run bru

  jcmp <- ~ x + z + Intercept(1)
  jfit <- bru(jcmp, lik1, lik2)

  bru_index(jfit, "norm")
  bru_index(jfit, "pois")

  # Compare the estimates

  p1 <- ggplot() +
    gg(fit1$summary.fixed, bar = TRUE) +
    ylim(0, 4) +
    ggtitle("Model 1")
  p2 <- ggplot() +
    gg(fit2$summary.fixed, bar = TRUE) +
    ylim(0, 4) +
    ggtitle("Model 2")
  pj <- ggplot() +
    gg(jfit$summary.fixed, bar = TRUE) +
    ylim(0, 4) +
    ggtitle("Joint model")

  multiplot(p1, p2, pj)
}

fbachl/inlabru documentation built on June 12, 2025, 2:09 p.m.