daisy

daisy provides estimation and bootstrap tools for Data-Adaptive Integration with Summary Data using generalized entropy balancing (GEB).

Main functions

• EB_est() — estimator (supports auto = TRUE model selection)
• EB_bootstrap_var() — bootstrap variance (no auto inside bootstrap; specify the model explicitly)

Installation

# using pak (recommended)
# install.packages("pak")
pak::pak("KMorikawaISU/daisy")

# or using remotes
# install.packages("remotes")
remotes::install_github("KMorikawaISU/daisy")

Quick start (paper-style toy data)

Simplest setting: [Internal] x1 ~ N(0,1), [External] x1 ~ N(0.5,1), x2 ~ Bern(0.5), **y | x ~ N(0.5*x1 − x2, 1)**,
Sample sizes: [internal] n = 200, [external] n_ext = 2000.
This chunk is **not executed** to keep the README light (eval = FALSE).

library(daisy)

## 1) Generate toy data
set.seed(1)
n     <- 200
n_ext <- 2000

x1_int <- rnorm(n, 0, 1)
x2_int <- rbinom(n, 1, 0.5)
y_int  <- rnorm(n, 0.5 * x1_int - x2_int, 1)

x1_ext <- rnorm(n_ext, 0.5, 1)
x2_ext <- rbinom(n_ext, 1, 0.5)
y_ext  <- rnorm(n_ext, 0.5 * x1_ext - x2_ext, 1)

## 2) Build inputs
dat_int <- data.frame(
  x1_int = x1_int,
  x2_int = x2_int,
  y_int  = y_int
)

# IMPORTANT: Do NOT include an intercept in MU_int / MU_ext.
MU_int <- cbind(x1_int, x1_int^2, x2_int)               # (x1, x1^2, x2)
MU_ext <- c(mean(x1_ext), mean(x1_ext^2), mean(x2_ext)) # external means for the same features

eta <- mean(y_ext)                                       # step-2 target (mean of outcome)

## 3-1) Estimation with auto model search
fit <- EB_est(
  dat_int, MU_int, MU_ext, eta,
  auto  = TRUE,
  r_set = c(0.01, 0.1, 0.5, 1),
  link  = "identity"
)

## Inspect results
fit$best_model
fit$result["est"]
fit$Entropy2

## Human-readable summary (point estimate + selected model + per-model Entropy2/D1/D2)
#summary(fit)


## 3-2) Estimation with specified model
#fit <- EB_est(
#  dat_int, MU_int, MU_ext, eta,
#  auto       = FALSE,
#  divergence = "LW",
#  r          = 0.1,
#  link       = "identity"
#)

## Human-readable summary (point estimate + selected model + per-model Entropy2/D1/D2)
summary(fit)


## 4) Bootstrap variance (no auto; fix the selected model)
#B1: Number of bootstrap repetitions for SigmaW_hat 
#B2: Number of bootstrap repetitions for theta_boot
bt <- EB_bootstrap_var(
  dat_int, MU_int, MU_ext, eta,
  n_ext = n_ext, B1 = 200, B2 = 500,
  w_type = TRUE,
  link   = "identity",
  divergence    = fit$best_model$divergence,
  r             = fit$best_model$r,
  external.boot = TRUE
)

bt$bootstrap_se
bt$ci_percentile_95

Notes

• First-step moments: MU_int / MU_ext must be provided without an intercept.
• Second step uses KL (fixed). The second-step weights are computed from the step-2 linear predictor (LMD2) obtained by KL calibration.