mlr.match: Thin wrapper around 'Match' function from 'Matching' package

In MatchLinReg: Combining Matching and Linear Regression for Causal Inference

Thin wrapper around Match function from Matching package

Description

Performs propensity score or Mahalanobis matching and return indexes of treatment and control groups.

Usage

mlr.match(tr, X, psm = TRUE, replace = F, caliper = Inf
  , verbose = TRUE)

Arguments

tr	Binary treatment indicator vector (1=treatment, 0=control), whose coefficient in the linear regression model is TE.
X	Covariates used in matching, either directly (Mahalanobis matching) or indirectly (propensity score).
psm	Boolean flag, indicating whether propensity score matching should be used (TRUE) or Mahalanobis matching (FALSE).
replace	Boolean flag, indicating whether matching must be done with or without replacement.
caliper	Size of caliper (standardized distance of two observations) used in matching. Treatment and control observations with standardized distance larger than caliper will not be considered as eligible pairs duing matching.
verbose	Boolean flag, indicating whether size of treatment and control groups before and after matching will be printed.

Details

For propensity score matching, linear predictors from logistic regression are used (rather than predicted probabilities).

Value

A vector of matched indexes, containing both treatment and control groups. Also, the following attributes are attached: 1) nt: size of treatment group, 2) nc: size of control group, 3) psm.reg: logistic regression object used in generating propensity scores (NA if psm is FALSE), 4) match.obj: matching object returned by Match function.

Author(s)

Alireza S. Mahani, Mansour T.A. Sharabiani

Examples

data(lalonde)

tr <- lalonde$treat
Z.i <- as.matrix(lalonde[, c("age", "educ", "black"
  , "hispan", "married", "nodegree", "re74", "re75")])
Z.o <- model.matrix(~ I(age^2) + I(educ^2) + I(re74^2) + I(re75^2) - 1, lalonde)

# propensity score matching on all covariates
idx <- mlr.match(tr = tr, X = cbind(Z.i, Z.o), caliper = 1.0, replace = FALSE)

# improvement in maximum single-covariate bias due to matching
bias.obj.before <- mlr.bias(tr = tr, Z.i = Z.i, Z.o = Z.o)
bias.before <- bias.obj.before$subspace$bias
dir <- bias.obj.before$subspace$dir
bias.after <- as.numeric(mlr.bias(tr = tr[idx]
  , Z.i = Z.i[idx, ], Z.o = dir[idx], gamma.o = 1.0)$single$bias)

# percentage bias-squared rediction
cat("normalized bias - before:", bias.before, "\n")
cat("normalized bias - after:", bias.after, "\n")
cat("percentage squared-bias reduction:"
  , (bias.before^2 - bias.after^2)/bias.before^2, "\n")

# matching with replacement
idx.wr <- mlr.match(tr = tr, X = cbind(Z.i, Z.o), caliper = 1.0
  , replace = TRUE)
bias.after.wr <- as.numeric(mlr.bias(tr = tr
  , Z.i = Z.i, Z.o = dir, gamma.o = 1.0, idx = idx.wr)$single$bias)
cat("normalized bias - after (with replacement):", bias.after.wr, "\n")

MatchLinReg documentation built on Aug. 30, 2022, 5:05 p.m.