apoorvalal / ebal

entropy balancing (v2)

ebal : Entropy Balancing

Updated version of the ebal R package for the entropy balancing method to compute method-of-moments based inverse propensity scores for causal inference and related problems.

Primary function is ebalance(W, X) to estimate the ATT (where W is the treatment vector).

New features: The main user-facing function ebalance optionally computes weights using automatic differentiation (via torch) when called with method="AutoDiff". This greatly improves stability (at the cost of marginally slower runtime) for small problems and greatly improves speed and stability for large problems. For detailed performance comparisons, see here.

Usage

library(ebal)
# lalonde observational data
load(url("https://github.com/apoorvalal/LalRUtils/raw/master/data/lalonde.psid.RData"))
yn = "re78"; wn = "treat"; xn = setdiff(colnames(lalonde.psid), c(yn, wn))

################################################################################
# naive estimate - this is badly biased relative to experimental benchmark of 1794
with(lalonde.psid, mean(re78[treat == 1]) - mean(re78[treat == 0]))
# -15204.7755

ebalance takes a treatment vector and covariates and returns a weight vector for control units.

# solving for weights
bal_wts = ebalance(lalonde.psid[['treat']], lalonde.psid[,xn], method = "AutoDiff")$w

which can then be fed into weighted.mean.

################################################################################
# balancing estimate - better
with(lalonde.psid, mean(re78[treat == 1]) - weighted.mean(re78[treat == 0], bal_wts))
# 2422.7965

or into a regression.

################################################################################
# augmented balancing estimate

# create weights vector in dataset for regression - uniform weights for treated units
lalonde.psid[['wt']] = 1/sum(lalonde.psid$treat)            # treat weights = 1/Nt
lalonde.psid[['wt']][lalonde.psid$treat == 0] = bal_wts     # ctrl weights = ebal weights

estimatr::lm_robust(re78 ~ . - wt, weights = wt, lalonde.psid) %>%
  summary %>% coefficients %>% .['treat',] %>% round(2)

# Estimate Std. Error    t value   Pr(>|t|)   CI Lower   CI Upper         DF
#  2422.79     750.36       3.23       0.00     953.00    3895.71    2663.00

installation

remotes::install_github("apoorvalal/ebal")

lower-level functions for usage with aggregate data, and other reweighting problems

Reweighting may be useful in a wide variety of settings (e.g. survey reweighting, experiment generalization, density ratio estimation). In many cases, you may only have summary statistics and not individual level covariates for the target sample.

In such cases, you may use eb or ebal_torch functions directly, where you supply a donor matrix X0, a target moment vector X1, and optionally supply base weights.

For example, if we only had control group averages in the above problem, we could solve for weights like so:

# target summary - this could come from a census
X1 = colMeans(lalonde.psid[lalonde.psid$treat == 1, xn])
# source matrix
X0 = lalonde.psid[lalonde.psid$treat == 0, xn] %>% as.matrix()
# weights that set weighted mean of X0 to X1
w = ebal_torch(X0, X1)$Weights.ebal

# same as above
with(lalonde.psid, mean(re78[treat == 1]) - weighted.mean(re78[treat == 0], w))
# 2422.7965