Extract temporal effects from a fitted model

Extract posterior distributions of temporal effects from a fitted tulpa model with temporal specification.

Usage

# S3 method for class 'tulpa_temporal_field_fit'
print(x, ...)

# S3 method for class 'tulpa_temporal_field'
print(x, ...)

temporal(
  object,
  component = "all",
  summary = FALSE,
  probs = c(0.025, 0.5, 0.975),
  ...
)

# S3 method for class 'tulpa_fit'
temporal(
  object,
  component = "all",
  summary = FALSE,
  probs = c(0.025, 0.5, 0.975),
  ...
)

# S3 method for class 'tulpa_temporal_posterior'
print(x, ...)

# S3 method for class 'tulpa_temporal_posterior'
summary(object, probs = c(0.025, 0.5, 0.975), ...)

# S3 method for class 'tulpa_temporal_summary'
print(x, ...)

Arguments

x

A temporal() field specification, the fit tulpa() returns for one, a tulpa_temporal_posterior or a tulpa_temporal_summary.

...

Ignored

object

A tulpa_fit object fitted with temporal argument

component

Which component to extract for multi-scale models: "all" (default), "trend", "seasonal", or "short_term".

summary

Logical; if TRUE, return summary statistics instead of full posterior draws.

probs

Quantiles to compute if summary = TRUE.

Value

A tulpa_temporal_posterior object

summary() on a tulpa_temporal_posterior returns a tulpa_temporal_summary data frame with one row per time point (per component for multi-scale fits), holding the posterior mean, SD and requested quantiles of the temporal effect.

Details

temporal() is overloaded. Given a fitted model it is the accessor described here. Given a one-sided formula (or a named formula = / structure = argument) it is instead the inline varying-coefficient field constructor used in a tulpa() model formula, the temporal mirror of spatial(): temporal(formula = ~ 1 + x || time, structure = "rw1") declares a smooth temporal level (the intercept column) plus a temporally varying slope on each covariate column. structure is one of "rw1" (default), "rw2", or "ar1"; only the double bar || (independent fields) is supported.

On a sampler fit the posterior is the field's own draws. On a nested-Laplace fit it is the Gaussian mixture the outer grid defines: each hyperparameter cell contributes its conditional mode and marginal variance, weighted by the cell's posterior weight, and summary() reports that mixture's exact mean, SD and quantiles. The draws of such a fit are sampled from the same mixture, each from its cell's joint Gaussian, so they carry the correlation between time points as well as each point's marginal, and a quantity derived from them (a trend, a difference between two times) has the right spread. A field so large that its per-cell covariance was not retained (more than 4 million values over the grid) is drawn with independent time points within each cell instead. The time column of the summary holds the time values the field was fitted on.

See also

temporal_multiscale(), temporal_rw1()

Examples

# \donttest{
set.seed(131)
df <- data.frame(year = 1:40, x = rnorm(40))
df$count <- rpois(40, exp(1 + 0.2 * df$x))

fit <- tulpa(
  count ~ x,
  data = df,
  family = "poisson",
  temporal = temporal_multiscale("year", trend = "rw2", seasonal = 12),
  mode = "exact",
  control = list(n_iter = 200L, n_warmup = 100L, seed = 1L)
)

# Extract all temporal effects
temp_post <- temporal(fit)
summary(temp_post)
# }