sensitivity_paths() reports robustness values for
the a-path, the b-path, and the indirect effect in the partial R-squared
metric of Cinelli and Hazlett (2020). One confounder enters both paths
through its partial R-squared with the mediator. For the indirect effect
it gives the robustness value under the joint-significance test (equal
to the smaller path value) and under the Sobel test (always smaller),
and their ratio. It accepts a robmedfit object or fitted
lm mediator and outcome models.
adjust_paths() returns the adjusted path and Sobel
statistics for a confounder with separate partial R-squared values for
the treatment, mediator, and outcome.
plot_paths() plots the adjusted statistics against
confounder strength.
effect_curvature() replaces
sensitivity_curvature() (now deprecated). It describes the
shape of an effect curve and is not a sensitivity analysis. It returns
curvature exactly zero when the curve is linear by construction
(previously it plotted rounding error), and gives percentile bootstrap
intervals for the curvature otherwise. plot_curvature() now
shows the effect curve and its curvature with intervals; the “local
fragility index” and normalised-curvature panels are removed.
The natural indirect and direct effects are now correct when the outcome model is not linear in the mediator (for example a logistic outcome). Previously the predicted mediator mean was plugged into the outcome model, which is exact only for linear outcomes. The outcome prediction is now averaged over the mediator’s distribution: exactly for a binary mediator and by a 15-point quantile rule for a Gaussian mediator.
cluster_var now makes the bootstrap resample whole
clusters. Previously it only stored cluster sizes and a fixed ICC of
0.05 that was not estimated; the ICC field is removed.
spline_df is deprecated and ignored (it never had an
effect). Put a spline of the treatment in the formulas instead,
e.g. splines::ns(X, df = 3).
augment() now returns the same weights used in the
balance diagnostics.
Documentation now states that effects are estimated by g-computation and that the inverse probability weights are used for balance diagnostics only.
Removed unused internal code.
Repository URLs updated to https://github.com/subirhait/RobustMediate.
Removed the mediation ITCV (“medITCV”) functions:
sensitivity_meditcv(),
sensitivity_meditcv_profile(), plot_meditcv(),
plot_meditcv_profile(), fragility_table(), and
their print methods, together with the "meditcv" and
"meditcv_profile" options of plot(). The
quantities they reported did not match Frank’s (2000) impact threshold
for a confounding variable, and the “minimum robustness” summary for the
indirect effect does not hold in general.
Removed the E-value x rho sensitivity surface,
tipping_table(), and the evalue_seq argument
of robustmediate(). The surface was not computed from
either the E-value or the sensitivity model of Imai et al. (2010).
glance() now reports rho_zero_NIE in place of
evalue_NIE and rho_min_NIE.
sensitivity_rho() implements the sensitivity
analysis of Imai, Keele, and Yamamoto (2010) for linear mediator and
outcome models: the indirect effect as a function of the correlation
rho between the two models’ errors, the value of
rho at which it is zero, and the equivalent product of
variance-explained parameters. robustmediate() runs it
automatically when the models are linear and stores the result in
$sensitivity.
plot_sensitivity() now plots the indirect effect
against rho.
plot_curvature() failed to draw because the
confidence ribbon was applied to panels without interval
bounds.
Added the missing tests/testthat.R runner, so the
test suite now runs during R CMD check. Fixed a balance
test that referred to a non-existent field.
Fixed bootstrap confidence interval computation. The previous
version used stats::update() to refit models on bootstrap
samples, which failed silently when bootstrap data frames had duplicate
row names produced by sampling with replacement. All confidence
intervals were returning NA. The fix replaces
stats::update() with direct stats::glm() calls
and resets row names on each bootstrap sample. All bootstrap replicates
now succeed and confidence intervals are correctly computed.
Fixed balance SMD computation: max_smd and
n_above now handle edge cases with near-constant covariates
without returning NA.
Added n_boot_valid diagnostic field to
fit$effects reporting the number of successful bootstrap
replicates.
Improved warning messages when bootstrap replicates fail.
Removed pkgdown URL from DESCRIPTION (was returning 404). Only the GitHub repository URL is now listed.