Skip to contents

Visualising fitted local estimators

Source: vignettes/visualising-fits.Rmd
visualising-fits.Rmd

RdistNN ships ggplot2-based views of fitted local estimators. ggplot2 (and, for dashboards, patchwork) are suggested dependencies: everything in this vignette degrades with an actionable message when they are missing, and every panel is a regular ggplot object that can be extended with +.

The running example is a CATE fit with oracle nuisance columns, evaluated along a profile in the first covariate.

library(RdistNN)

set.seed(42)
n <- 400
X <- cbind(X1 = runif(n, -2, 2), X2 = rnorm(n))
e <- plogis(0.4 * X[, "X1"])
tau <- 1 + sin(X[, "X1"])
m0 <- X[, "X1"] - 0.5 * X[, "X2"]
treat <- as.integer(runif(n) < e)
Y <- m0 + treat * tau + rnorm(n, sd = 0.5)

df <- data.frame(X, T = treat, Y = Y, e = e, m0 = m0, m1 = m0 + tau)
X0 <- cbind(X1 = seq(-1.5, 1.5, length.out = 11), X2 = 0)

fit <- cate_dnn(
  data = df, x_cols = c("X1", "X2"), y_col = "Y", t_col = "T",
  e_col = "e", m0_col = "m0", m1_col = "m1",
  X0 = X0, k = 1L, se = TRUE, jackknife = "fast",
  return = c("W", "dx")
)

The estimate panel

plot() (equivalently ggplot2::autoplot(), or plot_estimates() with more options) draws the profile with jackknife confidence bounds, a zero reference line, and the honesty caption: the median effective sample size across targets and the kernel order s with its provenance — here the default paper rule, since neither s nor tuning was supplied.

plot(fit)

Small non-profile target sets render as a forest layout with labels built from the stored coordinates, and style =/x_var = override the automatic choices.

Weight geometry

The weight panels show what the estimator averaged over. The effective kernel view is the most instructive: DNN weights as a data-driven kernel in distance space, with the rank-s radius marked.

plot_weights(fit, type = "kernel", targets = 6L)

plot_weights(fit, type = "cumulative", targets = c(1L, 6L, 11L))

Overlap diagnostics

For CATE fits, plot_overlap() shows whether both treatment arms carry weight near each target. With simplex weights the treated and control masses partition one, so the treated mass is the local treated share; shaded bands mark shares outside share_limits.

plot_overlap(fit, type = "mass")

The propensity view draws the weighted propensity distribution at selected targets when observation-level propensities are available (as here, through e_col), and a weighted q10–q90 summary otherwise.

plot_overlap(fit, type = "propensity", targets = c(1L, 6L, 11L))

Weight diagnostics facets

autoplot() on a diagnostics() frame tracks the effective sample size (with the degenerate-weights guide at 0.9 n), the maximum weight, and the top-5 weight share across targets.

ggplot2::autoplot(diagnostics(fit))

The dashboard

plot_dashboard() composes the panels with a metadata annotation. It requires patchwork; without it the panels are returned as a named list. Panels whose inputs are unavailable are substituted rather than failing the dashboard — a matrix-interface fit without stored treatment gets the diagnostics facets in place of the overlap panel.

Customising panels

Every builder returns a plain ggplot object; add layers, scales, or a different theme as usual.

plot_estimates(fit, level = 0.9) +
  ggplot2::labs(title = "CATE profile, 90% intervals") +
  ggplot2::theme_classic(base_size = 11)

For fully custom figures, the tidy frames behind the panels are exported: plot_data_estimates(), plot_data_weights(), and plot_data_diagnostics().