Fit one binary probability model for a support assessment

Fit a SuperLearner ensemble on an explicitly supplied training set and predict an explicitly supplied validation set. Use this for each conditional exposure regression and each observation regression in the study design.

Usage

margot_lmtp_support_probability(
  y,
  x,
  new_x,
  ids,
  learners,
  inner_folds,
  seed,
  method = "method.NNLS"
)

Arguments

y

Binary training responses.

x, new_x

Numeric data frames with identical columns for training and validation. Only variables available before the relevant event belong here.

ids

Unique training participant identifiers, passed to SuperLearner.

learners

Ordered SuperLearner library selected by the study.

inner_folds

Number of inner cross-validation folds, at least two.

seed

Explicit integer seed used before fitting a nonconstant response.

method

SuperLearner combination method; defaults to "method.NNLS".

Value

A list with state (fitted, constant, or unassessable), ensemble probability, learner_probability, a compact fit record, warnings, reason and seed. The compact record retains coefficients, cross-validated risks, error flags and fold sizes via the same helper as margot_lmtp_exposure_ratio(). It also retains the inner validation-row indices. Fitted learner objects are not returned.

Details

Empty and one-class training responses return constant predictions. An empty conditional set is usable only when its parent probability is zero; the calling conditional-law calculation must check this condition. Failed learners and invalid predictions give an unassessable result. Probabilities within 1e-12 outside [0, 1] are snapped to the nearest bound; larger departures remain invalid. Warnings are recorded and muffled, preserving the support-stage caller's behaviour. Invalid arguments raise an error before fitting; fitting failures return an unassessable record. The empty_training field explicitly marks an empty conditional set. The caller owns outer folds, checkpoints and worker/thread settings. This function sets the seed before a nonconstant fit and uses no future plan of its own. It does not inspect terminal outcomes or select a schedule.

Examples

margot_lmtp_support_probability(rep(1, 4), data.frame(x = 1:4),
  data.frame(x = 5:6), letters[1:4], "SL.mean", 2L, 42L)
#> $state
#> [1] "constant"
#> 
#> $probability
#> [1] 1 1
#> 
#> $learner_probability
#>      SL.mean_All
#> [1,]           1
#> [2,]           1
#> 
#> $fit
#> NULL
#> 
#> $reason
#> [1] "one observed class in the conditional training set"
#> 
#> $empty_training
#> [1] FALSE
#> 
#> $warnings
#> character(0)
#> 
#> $seed
#> [1] 42
#>