Trimmed inverse-probability weights (C backend)

The inverse-probability-of-treatment weights \(1/e\) for the treated and \(1/(1-e)\) for the untreated, with the propensity score clamped into [trim_lo, trim_hi] FIRST. Clamping matters: an untrimmed score near 0 or 1 produces a weight large enough for one observation to dominate the entire estimate.

Usage

core_ipw_weights(treat, propensity, trim_lo = 0.01, trim_hi = 0.99)

Arguments

treat

Numeric or logical treatment indicator; 1 / TRUE is treated.

propensity

Numeric vector of propensity scores, the same length as treat.

trim_lo, trim_hi

Clamp bounds for the score (defaults 0.01 and 0.99).

Value

A numeric vector of weights the length of treat.

Examples

treat <- c(1, 0, 1, 0)
e <- c(0.5, 0.25, 0.02, 0.9)

core_ipw_weights(treat, e)
#> [1]  2.000000  1.333333 50.000000 10.000000

# A balanced score gives weight 2 to either arm.
core_ipw_weights(c(1, 0), c(0.5, 0.5))
#> [1] 2 2

# Without trimming the third observation would carry weight 50; the
# default clamp holds it to 100 at the 0.01 floor, and a looser floor
# tames it further.
core_ipw_weights(treat, e, trim_lo = 0.10)
#> [1]  2.000000  1.333333 10.000000 10.000000

# Logical treatment indicators work too.
core_ipw_weights(c(TRUE, FALSE), c(0.4, 0.4))
#> [1] 2.500000 1.666667