Skip to contents

fitted() on a random forest, a GWR or a Bayesian GP fit returns one value per training row, in the order of object$data_sf. What the value is depends on the backend, and so does what summary() and model_metrics() without newdata report:

  • a random forest returns out-of-bag predictions: each row predicted by the trees that did not sample it (see "Fitted values are out-of-bag" in fit_rf_model()). A row that every tree sampled has no out-of-bag prediction and is NaN;

  • a GWR returns the values fit_gwr_model() stored on the fit: at each training location, the value of the local regression centred there. These are in-sample values (each observation sits inside its own window), so summary(), and model_metrics() without newdata, report an optimistic fit; use cv_gwr() for a spatially blocked estimate. An engine made elsewhere (a GWmodel result) is read from its SDF table instead;

  • a Bayesian GP returns the posterior expectation at the training locations, the column means of brms::posterior_epred(), in the units of the response (see "The response's centre and scale" in fit_bayesian_spatial_model()). These are in-sample values.

Usage

# S3 method for class 'gwr_fit'
fitted(object, ...)

# S3 method for class 'bayesian_fit'
fitted(object, ...)

# S3 method for class 'rf_fit'
fitted(object, ...)

Arguments

object

A fitted rf_fit, gwr_fit or bayesian_fit.

...

Not used.

Value

A numeric vector of length object$n. For a GWR, NA where the values cannot be read from the engine. For a Bayesian GP, a posterior that cannot be drawn is an error, as is a family with a probability per response category (ordinal, categorical), which has no single fitted value per row.

The Bayesian GP cache

brms::posterior_epred() takes time in proportion to the number of draws times the number of rows, and summary(), residuals(), model_metrics() and compare_models() each call fitted(), so the values are kept in an environment carried in object$info$.cache. The cache holds the expected values only, so predict(object, summary = "median") and predict(object, type = "predict") recompute from the posterior.

The entry records the n and a digest of the training data it was computed from, and is recomputed when either differs, so a copy given other data does not read the original's values. It is also tied to the sampling run: a refit or update() of the brmsfit recomputes. A brmsfit edited in place is what clear_fitted_cache() is for. A copy made with fit2 <- fit shares the environment, so clear_fitted_cache() on one empties the cache of both (the other recomputes), and identical() cannot tell two fits apart by their caches. The entry holds the values and a small identifier of the sampling run, so a fit saved with saveRDS() after fitted() is no larger for it.

Examples

library(sf)
set.seed(1)
pts <- st_as_sf(
  data.frame(x = 5e5 + runif(60, 0, 1000), y = 5e6 + runif(60, 0, 1000),
             a = rnorm(60)),
  coords = c("x", "y"), crs = 32632
)
pts$z <- 10 + 2 * pts$a + rnorm(60, 0, 0.3)
rf <- fit_rf_model(pts, "z", "a", num_trees = 50, seed = 1)
gw <- fit_gwr_model(pts, "z", "a", bandwidth = 30)
# Out-of-bag for the forest, in-sample for the GWR.
head(cbind(z = pts$z, rf = fitted(rf), gwr = fitted(gw)))
#>              z        rf       gwr
#> [1,] 14.651448 14.029639 14.619169
#> [2,] 10.324432  9.960532  9.882031
#> [3,] 11.315105 11.204729 11.347377
#> [4,] 10.002137  9.993318 10.070932
#> [5,]  8.483396  8.534352  8.563605
#> [6,] 10.591384 10.239645 10.411187