Skip to contents

Fit 2SLS using foundation-model (or proxy) embeddings as instruments

Source: R/causal_iv.R
causal_iv_from_embeddings.Rd

Convenience wrapper that:

  1. Takes a matrix of per-profile embeddings (rows = profiles, columns = embedding dims),

  2. Reduces them to n_pcs principal components,

  3. Attaches the PCs to data as new columns named PC_1, ...,

  4. Runs causal_iv_fit_2sls() with the PCs as instruments.

Usage

causal_iv_from_embeddings(
  data,
  embeddings,
  exposure,
  outcome,
  covariates = NULL,
  n_pcs = 5L
)

Arguments

data

Data frame with exposure, outcome and any covariates.

embeddings

Numeric matrix with nrow(data) rows (one per data row) and any number of columns (embedding dimensions).

exposure, outcome, covariates

See causal_iv_fit_2sls().

n_pcs

Integer; number of top principal components to keep as instruments. Default 5L.

Value

edaphos_causal_iv object (see causal_iv_fit_2sls()).

Details

Using the top n_pcs principal components instead of raw embedding dimensions keeps the instrument count manageable (avoiding the curse of dimensionality) and ensures the instruments are orthogonal (which simplifies the Sargan diagnostics). The default n_pcs = 5L yields a 4-over-identified model for a single-exposure query, enabling the Sargan J-test.