Skip to contents

Estimate a causal effect using DAG-guided backdoor adjustment

Source: R/causal_dag.R
causal_estimate_effect.Rd

Identifies a valid backdoor-adjustment set from the supplied DAG (unless one is provided manually) and then fits an adjusted outcome model conditional on that set. Two estimators are available:

Usage

causal_estimate_effect(
  data,
  dag,
  exposure,
  outcome,
  adjustment = NULL,
  effect = c("direct", "total"),
  type = c("minimal", "canonical", "all"),
  estimator = c("lm", "bart"),
  delta = NULL,
  bart_kwargs = list()
)

Arguments

data

Data frame with columns covering at least exposure, outcome, and the chosen adjustment set.

dag

A dagitty DAG.

exposure, outcome

Character column names.

adjustment

Optional character vector overriding the automatic adjustment set.

effect, type

Forwarded to causal_adjustment_set().

estimator

One of "lm" (default) or "bart" (requires dbarts).

delta

Numeric finite-difference step used by the BART estimator. Defaults to the interquartile range of exposure divided by two.

bart_kwargs

Optional named list of extra arguments forwarded to dbarts::bart() (e.g. ndpost, nskip, seed).

Value

A edaphos_causal_effect object with:

model

The fitted estimator (either an lm or a dbarts::bart object).

estimator

Character; "lm" or "bart".

adjustment

The adjustment set used.

effect

Numeric direct effect.

effect_ci

95 % CI (asymptotic for "lm", posterior quantile for "bart").

effect_naive

Coefficient from the unadjusted lm(outcome ~ exposure) for contrast.

Details

  • estimator = "lm" — closed-form linear regression \(Y = \beta_0 + \beta_{\text{exposure}}\,X + \sum_{z\in Z}\gamma_z z + \varepsilon\). The regression coefficient on exposure is the direct causal effect. Confidence intervals follow from OLS asymptotics.

  • estimator = "bart" — non-linear Bayesian Additive Regression Trees (Chipman, George & McCulloch 2010), via the dbarts Suggests dependency. The effect of exposure is computed as the average partial derivative \(\bar{\partial} = \frac{1}{n}\sum_i \bigl[\widehat{E}[Y\mid X=x_i+\delta, Z=z_i] - \widehat{E}[Y\mid X=x_i, Z=z_i]\bigr] / \delta\) averaged over the training data. A 95 % credible interval is recovered from the BART posterior draws.

References

Chipman, H. A., George, E. I., & McCulloch, R. E. (2010). BART: Bayesian Additive Regression Trees. Annals of Applied Statistics 4, 266-298.