ria.test 0.4.0
Breaking change
- In
ria.test(), replace the named arguments
outcome and mediators with out
and med, respectively. Their positions are unchanged.
Estimation and inference
- Sequential Riesz regression now pairs each observation with its own
previous-stage weight in every minibatch. This correction can change
estimates, standard errors, and p-values. Rerun analyses fitted with
earlier releases.
- Clustered standard errors now use within-cluster sums of influence
scores centered at the weighted mean. This corrects inference when
clusters differ in size or total weight and requires
ife
0.2.5 or later.
- The default sequential outcome regressions now combine LightGBM, the
sample mean, multivariate adaptive regression splines, and a
single-hidden-layer neural network with
decay = 0.001. Each
regression estimates nonnegative ensemble weights by cross-validation
(10 folds by default) and normalizes them to sum to one. The LightGBM
and MARS wrappers require mlr3extralearners from the
mlr-org R-universe.
- Regressions with a constant, finite continuous outcome use the mean
learner. Learner weights are aligned by name across folds.
- The defaults are now 5 cross-fitting folds and 2 L-prime folds.
Riesz networks now train for 10 epochs with a learning rate of 0.001 and
dropout of 0.2.
- The default Torch seed is now 1. Set
torch_seed = NULL
to leave Torch’s random-number stream alone. Use set.seed()
to control R-level randomness.
ria.test 0.3.0
- Use D for treatment, C for baseline covariates, and L for
post-treatment confounders throughout the implementation, examples, and
documentation. The outcome-observation indicator is named
observed internally. Rename the matched copy to L-prime and
its control to lprime_folds. This is a breaking rename: use
lprime_folds in place of zprime_folds. The
public arguments trt, pre, and
post are unchanged.
- L-prime permutations now bypass the LP when every treatment/baseline
profile in a fold occurs at least twice. Random cycles within identical
profiles give an exact zero-cost optimum without constructing distances
or an LP. This also handles all-identical profiles without dividing by
zero.
- Permutations return indices internally, and L is selected directly
instead of multiplying by a dense permutation matrix. Folds with
singleton profiles continue to use the full LP; solver status and
permutation validity are checked. Folds with fewer than two observations
produce an explicit error.
- The fast path is reproducible under
set.seed(), but
chooses tied optima and consumes random numbers differently from earlier
versions. Permutations and downstream estimates can therefore change for
the same seed.
- This bypasses the reported Rsymphony native crash for qualifying
folds; it does not fix or diagnose the underlying native defect. The
general LP path still depends on Rsymphony and retains its quadratic
size.
ria.test 0.2.1
Reproducibility
- Add
torch_seed to ria.test.control() to
control Torch neural-network initialization and dropout independently of
R’s random-number generator.
- Delay Torch device initialization until after R-randomized
estimation steps, so first-use Torch initialization cannot change
cross-fitting or Z-prime folds.
ria.test 0.2.0
Breaking changes
- Restrict estimation to natural effects, randomized interventional
analogues, and the TE - TE^R falsification test.
- Replace
effect with estimand = "test",
"natural", or "ria"; the test is now the
default.
- Align returned estimand names with the paper and remove nuisance
fits from the returned object.
ria.test 0.1.3
Bug fixes
- Synced fold construction and influence-function identifiers with
upstream fixes.
ria.test 0.1.2
Bug fixes
- Fixed a bug caused by upgrading to
S7 0.2.0.
ria.test 0.1.0
General
- Initial CRAN submission.
- Uses
ife for storing effect estimates.