predict: Predict from a glmmfields model

In glmmfields: Generalized Linear Mixed Models with Robust Random Fields for Spatiotemporal Modeling

Predict from a glmmfields model

Description

These functions extract posterior draws or credible intervals. The helper functions are named to match those in the rstanarm package and call the function predict() with appropriate argument values.

Usage

## S3 method for class 'glmmfields'
predictive_interval(object, ...)

## S3 method for class 'glmmfields'
posterior_linpred(object, ...)

## S3 method for class 'glmmfields'
posterior_predict(object, ...)

## S3 method for class 'glmmfields'
predict(
  object,
  newdata = NULL,
  estimate_method = c("median", "mean"),
  conf_level = 0.95,
  interval = c("confidence", "prediction"),
  type = c("link", "response"),
  return_mcmc = FALSE,
  offset = NULL,
  iter = "all",
  ...
)

Arguments

object	An object returned by glmmfields().
...	Ignored currently
newdata	Optionally, a data frame to predict on
estimate_method	Method for computing point estimate ("mean" or "median")
conf_level	Probability level for the credible intervals.
interval	Type of interval calculation. Same as for stats::predict.lm().
type	Whether the predictions are returned on "link" scale or "response" scale (Same as for stats::predict.glm()).
return_mcmc	Logical. Should the full MCMC draws be returned for the predictions?
offset	Optional offset vector to be used in prediction.
iter	Number of MCMC iterations to draw. Defaults to all.

Examples

library(ggplot2)

# simulate:
set.seed(1)
s <- sim_glmmfields(
  n_draws = 12, n_knots = 12, gp_theta = 2.5,
  gp_sigma = 0.2, sd_obs = 0.1
)

# fit:
# options(mc.cores = parallel::detectCores()) # for parallel processing
m <- glmmfields(y ~ 0,
  data = s$dat, time = "time",
  lat = "lat", lon = "lon",
  nknots = 12, iter = 800, chains = 1
)

# Predictions:
# Link scale credible intervals:
p <- predict(m, type = "link", interval = "confidence")
head(p)

# Prediction intervals on new observations (include observation error):
p <- predictive_interval(m)
head(p)

# Posterior prediction draws:
p <- posterior_predict(m, iter = 100)
dim(p) # rows are iterations and columns are data elements

# Draws from the linear predictor (not in link space):
p <- posterior_linpred(m, iter = 100)
dim(p) # rows are iterations and columns are data elements

# Use the `tidy` method to extract parameter estimates as a data frame:
head(tidy(m, conf.int = TRUE, conf.method = "HPDinterval"))

# Make predictions on a fine-scale spatial grid:
pred_grid <- expand.grid(
  lat = seq(min(s$dat$lat), max(s$dat$lat), length.out = 25),
  lon = seq(min(s$dat$lon), max(s$dat$lon), length.out = 25),
  time = unique(s$dat$time)
)
pred_grid$prediction <- predict(m,
  newdata = pred_grid, type = "response", iter = 100,
  estimate_method = "median", offset = rep(0, nrow(pred_grid))
)$estimate

ggplot(pred_grid, aes(lon, lat, fill = prediction)) +
  facet_wrap(~time) +
  geom_raster() +
  scale_fill_gradient2()

glmmfields documentation built on Oct. 21, 2023, 1:06 a.m.