Skip to contents

balancing

balancing calculates optimization-based balancing weights for causal inference. Rather than modeling the probability of exposure and then hoping the resulting weights balance the covariates, these methods make covariate balance the objective of a convex optimization problem and solve for the weights that achieve it directly. The package covers six methods (entropy balancing, inverse probability tilting, the covariate balancing propensity score, energy balancing, characteristic function distance balancing, and stable balancing weights) across binary, categorical, and continuous exposures and a range of estimands. A Rust core provides the numerical solvers.

Installation

You can install the development version of balancing from r-causal.r-universe.dev with:

install.packages(
  "balancing",
  repos = c("https://r-causal.r-universe.dev", getOption("repos"))
)

You can also install the development version of balancing from source from GitHub with:

# install.packages("pak")
pak::pak("r-causal/balancing")

Installing from source requires a Rust toolchain (rustc 1.88 or newer and Cargo); we recommend installing one with rustup.

Usage

You give balance() a data frame, name the exposure and covariates with tidyselect, choose a method and an estimand, and it returns a fitted object carrying the weights and a balance table.

library(balancing)

# Simulate data with two confounders and a binary exposure
set.seed(1)
n <- 500
x1 <- rnorm(n)
x2 <- rnorm(n)
z <- rbinom(n, 1, plogis(0.4 * x1 - 0.5 * x2))
y <- 1 + 0.9 * z + 0.6 * x1 - 0.4 * x2 + rnorm(n)
study <- data.frame(exposure = z, age = x1, score = x2, outcome = y)

# Fit entropy balancing weights for the ATT
fit <- balance(
  study,
  exposure,
  c(age, score),
  method = bw_entropy(),
  estimand = "att"
)

fit
#> 
#> ── Entropy balancing ───────────────────────────────────────────────────────────
#> Exposure: "exposure" (binary)
#> Estimand: "att" (focal level "1")
#> Observations: 500
#> Solver: converged in 3 iterations
#> Constraints: 2 terms (tolerance 0)
#> Largest imbalance: 0.0000 (standardized mean difference)

The printed summary reports the solver status and the largest imbalance the weights leave behind on the constraint terms.

The weights are a bw vector, a sibling of propensity::psw(). Pass them to a weighted outcome model to estimate the effect.

study$w <- weights(fit)
outcome_mod <- lm(outcome ~ exposure, data = study, weights = w)
coef(outcome_mod)[["exposure"]]
#> [1] 1.131677

For the estimating-equation methods (entropy balancing, inverse probability tilting, and the covariate balancing propensity score), pass the fit and the weighted outcome model to ipw() to get effect estimates with standard errors that account for having estimated the weights. Binary, categorical, and continuous exposures are supported; with a continuous exposure the reported effects are the exposure coefficients of a weighted marginal structural model.

ipw(fit, outcome_mod)
#> Inverse Probability Weight Estimator
#> Estimand: ATT 
#> Effects: marginal (population-averaged) 
#> 
#> Weight Estimator:
#>   Call: balance(.data = study, .exposure = exposure, .covariates = c(age, 
#>     score), method = bw_entropy(), estimand = "att") 
#> 
#> Outcome Model:
#>   Call: lm(formula = outcome ~ exposure, data = study, weights = w) 
#> 
#> Marginal estimates:
#>      estimate std.err      z ci.lower ci.upper conf.level   p.value    
#> diff  1.13168 0.12345 9.1673  0.88972   1.3736       0.95 < 2.2e-16 ***
#> ---
#> Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

How balancing relates to the other r-causal packages

balancing is part of the r-causal family and shares its design language with three sibling packages.

  • propensity calculates inverse probability weights from a fitted propensity score model and estimates effects with ipw(). balancing produces a different kind of weight, one that targets balance directly rather than through a modeled score, but it plugs into the same effect-estimation workflow: the weights share the causal_wts vocabulary, and ipw() dispatches on a balancing fit just as it does on a propensity model.
  • halfmoon assesses covariate balance. balancing reports the balance achieved on its own constraint terms, but a full balance-assessment workflow, including variables the weights were not asked to balance, lives in halfmoon. Extract the weights with weights() and pass them to halfmoon::check_balance() and halfmoon::plot_balance().
  • positively diagnoses positivity violations and extrapolation across binary, categorical, and continuous exposures. balancing redistributes influence across units to achieve balance, and positivity problems are exactly where that redistribution becomes extrapolation, showing up as extreme weights and a collapsing effective sample size, so checking positivity with positively complements weighting with balancing.

Learn more