Fit SDR from verified, previously estimated components

Use with margot_lmtp() through its lmtp_model_type argument. The caller supplies node density ratios, the realised fold map and the RNG procedure. No exposure or observation model is fitted and no additional cap is applied. The caller remains responsible for the provenance and scientific validity of the components. Input identities are checked before outcome fitting.

Usage

margot_lmtp_sdr_from_components(
  data,
  trt,
  outcome,
  shift,
  components,
  baseline,
  time_vary,
  learners_outcome,
  bounds,
  control = lmtp::lmtp_control(.trim = 1),
  progress_bar = function(...) invisible(NULL),
  learners_trt = NULL,
  allow_full_fits = NULL
)

Arguments

data

Analysis frame, including one continuous outcome and id.

trt

Ordered exposure column names.

outcome

Outcome column name.

shift

Registered shift function.

components

List containing frame (all non-outcome analysis columns), history, baseline, time_vary, trt, cens, density_ratios, folds, weights, seed, rng_kind, shift_source_sha256 and shifted_exposures.

baseline

Baseline covariate names, identical to components$baseline.

time_vary

List of covariates entering each node, identical to components$time_vary, including node order and within-node variable order.

learners_outcome

Outcome learner library.

bounds

Explicit outcome limits on the analysis scale.

control

An lmtp_control object with .trim = 1.

progress_bar

Optional native SDR progress callback.

learners_trt

Ignored; exposure components are already supplied.

allow_full_fits

NULL (default) or a single non-empty character string giving the reason for keeping full SuperLearner outcome fits. A control that sets .return_full_fits = TRUE is refused before fitting unless a reason is given, as in margot_lmtp(), which passes its own allow_full_fits on. A permitted reason is recorded in the fit's full_fits_reason element.

Value

An ordinary lmtp fit with supplied folds, histories, bounds and RNG provenance attached, and with the outcome regressions' warnings in warnings and the learners SuperLearner dropped in dropped_learners (both named "outcome"), as described in margot_lmtp().

Details

The current implementation supports continuous outcomes and requires a unique, nonmissing id column. It uses the internal interface of lmtp 1.5.4; other versions are rejected by the shared interface guard. At least two outer folds must be supplied, and the active future plan must provide exactly one worker per fold. This function does not change the future plan. Exposure waves may be nonconsecutive, and each node may include covariates from multiple earlier surveys; the caller declares when each covariate enters.