
Interpretable subgroup identification in randomized clinical trials, with post-selection inference for the effect the trial reports.
forestsearch identifies a subgroup defined by explicit covariate cuts, such as “no measured estrogen receptor and progesterone receptor at most 26 fmol/l”. It reports the trial’s standard analysis for that subgroup and for the patients outside it, corrected for the selection that produced the subgroup. Survival outcomes use the Cox hazard ratio. Binary, continuous and count outcomes use the coefficient of a generalized linear model: an odds ratio, relative risk, risk difference, mean difference or incidence-rate ratio.
The methods are described in León and Anderson (2026a, 2026b), and in León et al. (2024) for the original procedure (see Citation below).
maxk = 2) that meets minimum
size and event counts. The candidates come from the covariates alone; no
outcome model decides them.p*. The rate is the probability that both halves of
a random half-split clear the consistency threshold, and it is evaluated
in closed form from one fit
(consistency_method = "resample", the default).sg_focus = "effMaxSG" with
effect_neighborhood = 0.20 takes the largest subgroup whose
effect is within 20% of the strongest. The search can also return no
subgroup.mr_inference = TRUE,
multiplier resampling re-applies the selection rule to jointly perturbed
candidate effects, without refitting. It reports a de-biased estimate
for the subgroup and for its complement, a one-sided selection-adjusted
bound for each, and a Bonferroni pair for a joint claim on both.fs_declaration_calibration() estimates how often the
screen, as run, would declare a subgroup if every patient benefited by a
clinically specified amount. It also gives the cutoff that would hold
that rate at a nominal level.Model-based identifiers can propose the candidates instead: DINA
(subgroup_method = "dina") and causal forests
(subgroup_method = "grf"). Their regions are re-ranked on
the same coefficient and receive the same inference, conditional on the
candidate family they propose. The full bootstrap correction and the
cross-validation of León et al. (2024) remain available.
Bounds are read by where they lie against the trial’s own effect and against clinically meaningful effect sizes, not as tests at a hazard ratio of one.
From CRAN:
install.packages("forestsearch")The development version from GitHub:
# install.packages("pak")
pak::pak("larry-leon/forestsearch")This is the forest-search analysis of the German Breast Cancer Study Group trial in Table 1 of León and Anderson (2026b). The article Interpretable subgroup identification with forestsearch, listed under Articles on the package website, runs it in full, with the two model-based identifiers, and reads every number.
library(forestsearch)
library(survival)
df <- within(gbsg, {
id <- seq_len(nrow(gbsg))
time_months <- rfstime / 30.4375
grade3 <- ifelse(grade == "3", 1, 0)
})
fit <- forestsearch(
df.analysis = df,
outcome.name = "time_months", event.name = "status",
treat.name = "hormon", id.name = "id",
confounders.name = c("age", "meno", "size", "grade3", "nodes", "pgr", "er"),
is.RCT = TRUE, est.scale = "hr", seedit = 8316951,
# the candidate family
conf.cont_jcuts = list(er = 10, pgr = 10), conf_force = c("er <= 0", "pgr <= 0"),
n.min = 60, d0.min = 10, d1.min = 10, maxk = 2,
# the thresholds and the selection rule
hr.threshold = 1.00, hr.consistency = 1.00, pconsistency.threshold = 0.90,
stop_threshold = NULL,
sg_focus = "effMaxSG", selection_rule = "neighborhood", effect_neighborhood = 0.20,
# selection-adjusted inference
mr_inference = TRUE,
mr_inference_args = list(ci_method = "field", draws = 5000L, include_complement = TRUE,
field_scale_complement = "selected", field_decompose = TRUE,
keep_declaration_field = TRUE,
declaration_c0 = c(0.70, 0.75, 0.80, 0.85)),
parallel_args = list(plan = "sequential")
)
print(fit)
summary(fit)
# Would a screen protecting a uniform benefit (HR 0.70 to 0.85) have declared it?
fs_declaration_calibration(fit, alpha = 0.10, c0 = c(0.70, 0.75, 0.80, 0.85))| Function | Purpose |
|---|---|
forestsearch() |
Candidate enumeration, screening, selection and (with
mr_inference = TRUE) selection-adjusted inference, in one
call |
fs_mr_inference() |
Multiplier-resampling inference on a supplied candidate family |
fs_declaration_calibration() |
Family-wise declaration rate of the screen at a protected level, and the calibrated cutoff |
mr_estimates_table() |
The multiplier-resampling estimates in the layout of the bootstrap tables |
guohe_algorithm3() |
The de-biased estimate and one-sided bootstrap bound of Guo and He (2021), for comparison |
forestsearch_bootstrap_dofuture(),
summarize_bootstrap_results() |
The full bootstrap bias correction of León et al. (2024) |
forestsearch_Kfold(),
forestsearch_tenfold(),
cv_metrics_tables() |
Cross-validation of the identification |
plot_subgroup_results_forestplot(),
plot_sg_weighted_km() |
Forest plots and weighted Kaplan-Meier curves |
generate_aft_dgm_flex(),
simulate_from_dgm(),
run_simulation_analysis() |
Data-generating mechanisms and the simulation framework |
mrct_region_sims(),
create_dgm_for_mrct() |
Regional consistency in multi-regional trials |
?forestsearch documents every argument, with examples
on this trial and on the binary and continuous ACTG 175 endpoints.If you use forestsearch in your work, please cite:
@article{Leon2024,
author = {Le{\'o}n, Larry F. and Jemielita, Thomas and Guo, Zifang
and Marceau West, Rachel and Anderson, Keaven M.},
title = {Exploratory subgroup identification in the heterogeneous
{C}ox model: A relatively simple procedure},
journal = {Statistics in Medicine},
year = {2024},
doi = {10.1002/sim.10163}
}
@misc{LeonAnderson2026a,
author = {Le{\'o}n, Larry F. and Anderson, Keaven M.},
title = {Inference for standard trial estimands after data-driven
subgroup discovery},
year = {2026},
eprint = {2609.38361},
archivePrefix = {arXiv},
primaryClass = {stat.ME},
doi = {10.48550/arXiv.2609.38361}
}
@misc{LeonAnderson2026b,
author = {Le{\'o}n, Larry F. and Anderson, Keaven M.},
title = {Interpretable subgroup identification across survival and
generalized-linear outcomes},
year = {2026},
eprint = {2610.28379},
archivePrefix = {arXiv},
primaryClass = {stat.ME},
doi = {10.48550/arXiv.2610.28379}
}MIT © Larry F. Leon