Skip to contents

Calculate Propensity Score Weights for Causal Inference

Source: R/weights.R
wt_ate.Rd

This family of functions computes propensity score weights for various causal estimands:

  • ATE (Average Treatment Effect)

  • ATT (Average Treatment Effect on the Treated)

  • ATU (Average Treatment Effect on the Untreated, sometimes called the ATC, where the "C" stands for "control")

  • ATM (Average Treatment Effect for the Evenly Matchable)

  • ATO (Average Treatment Effect for the Overlap population)

    The propensity score can be provided as a numeric vector of predicted probabilities or as a data.frame where each column represents the predicted probability for a level of the exposure. They can also be propensity score objects created by ps_trim(), ps_refit(), or ps_trunc()

    The returned weights are encapsulated in a psw object, which is a numeric vector with additional attributes that record the estimand, and whether the weights have been stabilized, trimmed, or truncated.

Usage

wt_ate(
  .propensity,
  .exposure,
  .sigma = NULL,
  exposure_type = c("auto", "binary", "categorical", "continuous"),
  .treated = NULL,
  .untreated = NULL,
  stabilize = FALSE,
  stabilization_score = NULL,
  ...
)

wt_att(
  .propensity,
  .exposure,
  exposure_type = c("auto", "binary", "categorical", "continuous"),
  .treated = NULL,
  .untreated = NULL,
  ...
)

wt_atu(
  .propensity,
  .exposure,
  exposure_type = c("auto", "binary", "categorical", "continuous"),
  .treated = NULL,
  .untreated = NULL,
  ...
)

wt_atm(
  .propensity,
  .exposure,
  exposure_type = c("auto", "binary", "categorical", "continuous"),
  .treated = NULL,
  .untreated = NULL,
  ...
)

wt_ato(
  .propensity,
  .exposure,
  exposure_type = c("auto", "binary", "categorical", "continuous"),
  .treated = NULL,
  .untreated = NULL,
  ...
)

Arguments

.propensity

Either a numeric vector of predicted probabilities or a data.frame where each column corresponds to a level of the exposure.

.exposure

The exposure variable. For binary exposures, a vector of 0s and 1s; for continuous exposures, a numeric vector.

.sigma

For continuous exposures, a numeric vector of standard errors used with dnorm(). For example, this can be derived from the influence measures of a model (e.g., influence(model)$sigma).

exposure_type

Character string specifying the type of exposure. Options are "auto", "binary", "categorical", and "continuous". Defaults to "auto", which detects the type automatically.

.treated

The value representing the treatment group. If not provided, it is automatically detected.

.untreated

The value representing the control group. If not provided, it is automatically detected.

stabilize

Logical indicating whether to stabilize the weights. For ATE weights, stabilization multiplies the weight by either the mean of .exposure or the supplied stabilization_score.

stabilization_score

Optional numeric value for stabilizing the weights (e.g., a predicted value from a regression model without predictors). Only used when stabilize is TRUE.

...

Reserved for future expansion. Not currently used.

Value

A psw object (a numeric vector) with additional attributes:

  • estimand: A description of the estimand (e.g., "ate", "att").

  • stabilized: A logical flag indicating if stabilization was applied.

  • trimmed: A logical flag indicating if the weights are based on trimmed propensity scores.

  • truncated: A logical flag indicating if the weights are based on truncated propensity scores.

Details

Exposure Types

The functions support different types of exposures:

  • binary: For dichotomous treatments (e.g. 0/1).

  • continuous: For numeric exposures. Here, weights are calculated via the normal density using dnorm().

  • categorical: Currently not supported (an error will be raised).

  • auto: Automatically detects the exposure type based on .exposure.

Stabilization

For ATE weights, stabilization can improve the performance of the estimator by reducing variance. When stabilize is TRUE and no stabilization_score is provided, the weights are multiplied by the mean of .exposure. Alternatively, if a stabilization_score is provided, it is used as the multiplier.

Trimmed and Truncated Weights

In addition to the standard weight functions, versions exist for trimmed and truncated propensity score weights created by ps_trim(), ps_trunc(), and ps_refit(). These variants calculate the weights using modified propensity scores (trimmed or truncated) and update the estimand attribute accordingly.

The main functions (wt_ate, wt_att, wt_atu, wt_atm, and wt_ato) dispatch on the class of .propensity. For binary exposures, the weights are computed using inverse probability formulas. For continuous exposures (supported only for ATE), weights are computed as the inverse of the density function evaluated at the observed exposure.

See also

  • psw() for details on the structure of the returned weight objects.

Examples

## ATE Weights with a Binary Exposure

# Simulate a binary treatment and corresponding propensity scores
propensity_scores <- c(0.2, 0.7, 0.5, 0.8)
treatment <- c(0, 1, 0, 1)

# Compute ATE weights (unstabilized)
weights_ate <- wt_ate(propensity_scores, .exposure = treatment)
#> ℹ Treating `.exposure` as binary
weights_ate
#> <psw{estimand = ate}[4]>
#> [1] 1.250000 1.428571 2.000000 1.250000

# Compute ATE weights with stabilization using the mean of the exposure
weights_ate_stab <- wt_ate(propensity_scores, .exposure = treatment, stabilize = TRUE)
#> ℹ Treating `.exposure` as binary
weights_ate_stab
#> <psw{estimand = ate}[4]>
#> [1] 0.6250000 0.7142857 1.0000000 0.6250000

## ATT Weights for a Binary Exposure

propensity_scores <- c(0.3, 0.6, 0.4, 0.7)
treatment <- c(1, 1, 0, 0)

# Compute ATT weights
weights_att <- wt_att(propensity_scores, .exposure = treatment)
#> ℹ Treating `.exposure` as binary
weights_att
#> <psw{estimand = att}[4]>
#> [1] 1.0000000 1.0000000 0.6666667 2.3333333