Predict from a plqr_result object at new targets (experimental)

Predict from a plqr_result object at new targets (experimental)

Usage

# S3 method for class 'plqr_result'
predict(
  object,
  newdata = NULL,
  X0 = NULL,
  x0 = NULL,
  se.fit = FALSE,
  interval = c("none", "confidence"),
  level = 0.95,
  ...
)

Arguments

object: A plqr_result object from plqr_dnn().
newdata: Optional matrix/data frame of target covariates.
X0: Optional numeric matrix of targets (rows).
x0: Optional numeric vector for a single target.
se.fit: Logical; if TRUE, also return direct coefficient standard errors. The fitted object must retain density values from density or density_col.
interval: One of "none" or "confidence".
level: Confidence level for interval output.
...: Not used.

Value

With interval = "none" and se.fit = FALSE, a coefficient matrix with one row per PLQR coefficient and one column per target. With se.fit = TRUE, a list with fit and se.fit matrices. With interval = "confidence", a list with fit, lwr, and upr matrices, and also se.fit when requested. The coefficient source is attached to the fit as attr(., "theta_source").

Details

The returned fit matrix repeats the supplied theta_eval (meta$theta_source = "supplied"). Target-specific variation enters only through the DNN weights and the resulting direct-sandwich standard errors; no target-specific PLQR root is solved.

Coefficient-level standard errors and intervals require the direct PLQR sandwich (fit with density or density_col); a componentwise moment-level jackknife SE on a fit without a density input is not by itself a coefficient-level interval — see markdown_notes/design/15_plqr_variance_estimation_rationale.md. Prediction standard errors always use the analytic sandwich meat, regardless of the fit's theta_meat choice, because the fitted object does not retain the outcome data needed to re-run the jackknife at new targets.

Lifecycle

Experimental. PLQR functionality is outside the CATE and regression workflow focus of the initial public release. Its API and inference interfaces may change in later releases.

Examples

set.seed(3)
demo <- plqr_demo_data(n = 80, p = 2, p_d = 2, tau = 0.5, seed = 3)
fit <- plqr_dnn(
  data = demo$data,
  x_cols = c("X1", "X2"),
  D_cols = c("D1", "D2"),
  g_col = "g_oracle",
  mD_cols = c("mD1", "mD2"),
  y_col = "Y",
  theta_eval = demo$truth$theta_true,
  tau = 0.5,
  x0 = c(0, 0),
  method = "exact",
  k = 1L, s = 5L,
  se = FALSE
)
predict(fit, x0 = c(0.2, -0.1))
#>    target_1
#> D1      0.5
#> D2      0.5
#> attr(,"theta_source")
#> [1] "supplied"