Causal Active Learning: query the next sample(s) that most reduce the uncertainty of a targeted causal effect

Closes the loop between Pillar 1 (causal identification via backdoor adjustment) and Pillar 5 (autonomous active learning): instead of choosing candidates by marginal-Y uncertainty (classical AL), we choose by expected shrinkage of Var(beta_hat_{X->Y}).

Usage

al_query_causal(
  data,
  pool,
  dag,
  exposure,
  outcome,
  adjustment = NULL,
  n_select = 5L,
  strategy = c("leverage", "bootstrap"),
  B = 200L,
  seed = NULL
)

Arguments

data

Data frame with the exposure, outcome and adjustment columns (the current labelled set).

pool

Data frame with the same columns as data; the unlabelled candidate pool to query.

dag

A dagitty DAG.

exposure, outcome

Character; column names of the causal query.

adjustment

Optional character vector of adjustment-set columns. Inferred from dag via causal_adjustment_set() when NULL.

n_select

Integer; how many candidates to return. Default 5L.

strategy

One of "leverage" (hat-matrix diagonal, closed-form, fast) or "bootstrap" (re-bootstrap the effect for each candidate; slow but exact under the linear-OLS estimator).

B

Bootstrap replications when strategy = "bootstrap".

seed

RNG seed for bootstrap reproducibility.

Value

A data frame with one row per candidate, columns pool_index, leverage, expected_var_reduction, sorted by descending expected reduction. The top-n_select rows are the recommended next samples.