ria.test implements the empirical test introduced by Yu, Ge, and Elwert (2025)
for detecting when randomized interventional analogues should not be
interpreted as natural mediation effects. The test estimates
TE - TE^R, the difference between the total effect and its
randomized interventional analogue. Rejecting TE - TE^R = 0
falsifies the composite null that the natural indirect and direct
effects equal their randomized interventional analogues. The procedure
remains valid in settings where the natural effects themselves are not
identified, and |TE - TE^R| provides a lower bound on the
total divergence between the natural and randomized interventional
decompositions. The estimation uses a Riesz-regression approach.
ria.test is derived from the crumble package with various adaptations. {ria.test} uniquely enables statistical inference for the estimated TE − TER. Separately, {ria.test} also makes improvements in terms of reproducibility, efficiency, and the handling of edge cases.
install.packages("ria.test")The default ensemble uses the LightGBM and MARS wrappers from
mlr3extralearners. Install these once before running the
example below. If you use learners = "glm", you can skip
this step.
install.packages("mlr3extralearners",
repos = c("https://mlr-org.r-universe.dev", "https://cloud.r-project.org"))library(ria.test)
set.seed(123)
n <- 500
C <- rnorm(n)
D <- rbinom(n, 1, plogis(C))
L <- rnorm(n, D + C)
M <- rnorm(n, D + L + C)
Y <- rnorm(n, D + L + M + C)
dat <- data.frame(C, D, L, M, Y)
test_fit <- ria.test(
data = dat,
trt = "D",
out = "Y",
med = "M",
pre = "C",
post = "L",
d0 = \(data, trt) rep(0, nrow(data)),
d1 = \(data, trt) rep(1, nrow(data)),
control = ria.test.control(torch_seed = 123L)
)
tidy(test_fit)