benchmark_parameters: Helper function for benchmarking for sensitivity parameters

In senseweight: Sensitivity Analysis for Weighted Estimators

Helper function for benchmarking for sensitivity parameters

Description

Helper function that returns parameter estimates for an omitted variable with specified relative confounding strength to an observed covariate (or set of covariates)

Usage

benchmark_parameters(
  weights,
  weights_benchmark,
  k_sigma = 1,
  k_rho = 1,
  Y,
  Z,
  sigma2,
  estimand = "ATT"
)

Arguments

weights	Vector of estimated weights
weights_benchmark	Vector of estimated weights, omitting the covariate (or set of covariates) being used to benchmark
k_sigma	Relative ability of omitted confounder to explain variation in the true weights. If k_sigma > 1, then we expect the omitted confounder to explain more variation in the true weights than the benchmarked covariate(s). If k_sigma < 1, then we expect the omitted confounder to explain less of the variation in the true weights than the benchmarked covariate(s). Default is set to 1.
k_rho	Relative correlation of omitted confounder with the outcome. If k_rho > 1, then we expect the omitted confounder to be more correlated with the outcome than the benchmarked covariate(s). If k_rho < 1, then we expect the omitted confounder to be less correlated with the outcome than the benchmarked covariate(s). Default is set to 1.
Y	Vector of outcomes
Z	Vector of treatment assignment (Only necessary if estimand = "PATE" in order to estimate covariances)
sigma2	Estimated variance of the outcome (i.e., stats::var(Y) for obervational setting; stats::var(tau) for generalization setting)
estimand	String specifying the estimand of interest. Valid inputs are "PATE", "Augmented", "ATT", or "Survey".

Value

data.frame containing estimated parameter values for a confounder with specified relative confounder strength to an observed covariate (or set of covariates), as well as the estimated bias from such an omitted confounder.

Examples

data(jtpa_women)
site_name <- "NE"
df_site <- jtpa_women[which(jtpa_women$site == site_name), ]
df_else <- jtpa_women[which(jtpa_women$site != site_name), ]

# Estimate unweighted estimator:
model_dim <- estimatr::lm_robust(Y ~ T, data = df_site)
PATE <- coef(lm(Y ~ T, data = df_else))[2]
DiM <- coef(model_dim)[2]
# Generate weights using observed covariates:
df_all <- jtpa_women
df_all$S <- ifelse(jtpa_women$site == "NE", 1, 0)
model_ps <- WeightIt::weightit(
  (1 - S) ~ . - site - T - Y, 
  data = df_all, method = "ebal", estimand = "ATT"
)
weights <- model_ps$weights[df_all$S == 1]
# Estimate IPW model:
model_ipw <- estimatr::lm_robust(Y ~ T, data = df_site, weights = weights)
ipw <- coef(model_ipw)[2]
# Estimate bound for var(tau):
vartau <- var(df_site$Y[df_site$T == 1]) - var(df_site$Y[df_site$T == 0])
RV <- robustness_value(estimate = ipw, b_star = 0, sigma2 = vartau, weights = weights)

# Select weighting variables:
weighting_vars <- names(df_all)[which(!names(df_all) %in% c("site", "S", "Y", "T"))]

# Run benchmarking:
df_benchmark <- run_benchmarking(
  weighting_vars = weighting_vars,
  data = df_all[, -1],
  treatment = "T", outcome = "Y", selection = "S",
  estimate = ipw,
  RV = RV, sigma2 = vartau,
  estimand = "PATE"
)

print(df_benchmark)

senseweight documentation built on Aug. 23, 2025, 1:11 a.m.