Skip to contents

Overlap Weights and Trimming for Propensity Score Analysis

overlap_weights.Rd

Computes overlap weights (also called average overlap weights, ATO weights) or trimmed inverse probability weights. Overlap weights target the population with the most covariate overlap and have the desirable property of minimizing the asymptotic variance among all balancing weights.

Usage

overlap_weights(
  data,
  treatment,
  covariates,
  ps_formula = NULL,
  weight_type = c("overlap", "ipw", "att", "trim"),
  trim_threshold = 0.1,
  normalize = TRUE,
  plot = TRUE
)

Arguments

data

A data frame.

treatment

Character. Name of the binary treatment indicator (0/1).

covariates

Character vector. Covariates for the propensity score model.

ps_formula

Formula or NULL. Custom propensity score formula. If NULL, uses a main-effects logistic regression on covariates.

weight_type

Character. Type of weight: "overlap" (default, ATO), "ipw" (ATE weights), "att" (ATT weights), "trim" (trimmed IPW).

trim_threshold

Numeric. Propensity scores below this or above 1 - trim_threshold are trimmed (excluded). Only used when weight_type = "trim". Default 0.1.

normalize

Logical. Whether to normalize weights to sum to 1 within each treatment arm. Default TRUE.

plot

Logical. If TRUE, produce a propensity score overlap plot. Default TRUE.

Value

A list with:

weights

Numeric vector of length nrow(data).

ps

Numeric vector of propensity scores.

ess_treated

Effective sample size in the treated arm.

ess_control

Effective sample size in the control arm.

n_trimmed

Number of trimmed observations (only for "trim").

plot

ggplot2 overlap plot.

Details

Overlap weights for unit \(i\) are: $$w_i = D_i (1 - e_i) + (1 - D_i) e_i$$ where \(e_i = P(D=1|X_i)\). These automatically down-weight units near the boundaries of the propensity score distribution.

References

Li, F., Morgan, K. L., & Zaslavsky, A. M. (2018). Balancing covariates via propensity score weighting. Journal of the American Statistical Association, 113(521), 390–400.

Crump, R. K., Hotz, V. J., Imbens, G. W., & Mitnik, O. A. (2009). Dealing with limited overlap in estimation of average treatment effects. Biometrika, 96(1), 187–199.

Examples

data(lalonde, package = "MatchIt")
ow <- overlap_weights(
  data       = lalonde,
  treatment  = "treat",
  covariates = c("age", "educ", "re74", "re75")
)
ow$ess_treated
#> [1] 181.5
ow$plot