Complete Balance Assessment and Treatment Effect Estimation

A unified function for covariate balance assessment and treatment effect estimation. Combines a formal balance test (via classification permutation test), visual diagnostics (propensity score distributions), and treatment effect estimates using both difference-in-means and augmented inverse propensity weighting (AIPW).

Usage

balance(
  Y = NULL,
  W,
  X,
  alpha = 0.05,
  perm.N = 1000,
  class.method = "ferns",
  seed = 1995,
  fastcpt.args = list()
)

# S3 method for class 'balance'
print(x, ...)

# S3 method for class 'balance'
summary(object, ...)

# S3 method for class 'balance'
plot(x, which = "all", combined = TRUE, breaks = 15, ...)

Arguments

Y

Outcome vector (numeric) or NULL. If NULL, treatment effect estimation (and the treatment effect plot) is skipped.

W

Treatment assignment vector (binary: 0/1 or logical).

X

Pre-treatment covariate matrix or data frame.

alpha

Significance level for balance test. Default is 0.05.

perm.N

Number of permutations for the balance test. Default is 1000.

class.method

Classification method for balance test. Default is "ferns".

seed

Random seed for reproducibility. Default is 1995.

fastcpt.args

A named list of additional arguments to pass to fastcpt. For example, fastcpt.args = list(parallel = TRUE, leaveout = 0.2).

x

A balance result object.

...

Additional arguments (currently unused).

object

A balance result object (for summary method).

which

Character vector specifying which plots to create. Options are "pscores", "null_dist", "effects", or "all".

combined

Logical. If TRUE, displays all three plots in a combined panel. Default is TRUE.

breaks

Number of breaks for histograms. Default is 15.

Value

A list of class "balance" containing:

balance_test

Results from fastcpt including p-value and propensity scores.

dim

Difference-in-means estimate with standard error and confidence interval (only if Y is provided).

aipw

Doubly robust estimate from causal forest (with propensity weighting) with standard error and CI (only if Y is provided).

aipw_const

Outcome-adjusted estimate from causal forest (no propensity weighting) with SE and CI (only if Y is provided).

passed

Logical indicating whether the balance test passed.

alpha

The significance level used.

cf

The fitted causal_forest object for advanced users (only if Y is provided).

Examples

if (FALSE) { # \dontrun{
# Generate example data
n <- 500
p <- 10
X <- matrix(rnorm(n * p), n, p)
W <- rbinom(n, 1, 0.5)
Y <- W * 0.5 + X[,1] * 0.3 + rnorm(n)

# Run complete balance assessment
result <- balance(Y, W, X)
result
summary(result)
plot(result)
} # }