The hardware and bandwidth for this mirror is donated by METANET, the Webhosting and Full Service-Cloud Provider.
If you wish to report a bug, or if you are interested in having us mirror your free-software or open-source project, please feel free to contact us at mirror[@]metanet.ch.
MetaHunt is an R package for privacy-preserving meta-analysis of function-valued quantities (e.g. regression functions, conditional average treatment effect functions) across heterogeneous studies (Shi, Imai, & Zhang 2024, arXiv:2604.23847).
Documentation: https://wshi18.github.io/MetaHunt/
The package implements the methodology of
Shi, W., Imai, K., and Zhang, Y. Privacy-preserving Meta-analysis through Low-Rank Basis Hunting.
Given aggregate-level estimates of a function of interest from several source studies, together with study-level covariates, MetaHunt recovers a small set of latent basis functions that span all studies under a low-rank assumption, models how the mixing weights depend on study-level covariates via Dirichlet regression, and predicts the corresponding function for a new target population with valid conformal prediction intervals. Individual-level data from source studies are not required, which makes the procedure suitable for federated and multi-site settings.
The package also exports minmax_regret(), a
covariate-free baseline based on the worst-case-regret aggregator of
Zhang, Huang, and Imai (2024, arXiv:2412.11136). It is
useful when the number of source studies is small or the study-level
covariates are uninformative.
# install.packages("remotes")
remotes::install_github("WShi18/MetaHunt")library(MetaHunt)
set.seed(1)
# Simulate m studies, each represented by its function on a shared grid.
G <- 30; m <- 60
x <- seq(0, 1, length.out = G)
basis <- rbind(sin(pi * x), cos(pi * x), x)
W <- data.frame(w1 = rnorm(m), w2 = rnorm(m))
eta <- as.matrix(W) %*% cbind(c(1, -0.5), c(-0.4, 1), c(0, 0))
pi_true <- exp(eta) / rowSums(exp(eta))
F_hat <- pi_true %*% basis + matrix(rnorm(m * G, sd = 0.05), m, G)
# Fit and predict for a new study with W_new = (0, 0).
fit <- metahunt(F_hat, W, K = 3, dfspa_args = list(denoise = FALSE))
W_new <- data.frame(w1 = 0, w2 = 0)
pred <- predict(fit, newdata = W_new, wrapper = mean)
# 95% conformal interval for the same target.
ci <- split_conformal(F_hat, W, W_new, K = 3,
wrapper = mean, alpha = 0.05, seed = 1,
dfspa_args = list(denoise = FALSE))
ci$prediction
c(ci$lower, ci$upper)See vignette("metahunt-intro") for a full walkthrough
including data preparation from fitted model objects, rank selection,
conformal intervals (split / cross / pre-fit), and the minimax-regret
baseline.
metahunt() + predict(),
plot(), summary() — end-to-end fitter.split_conformal(),
cross_conformal(),
conformal_from_fit() — prediction intervals.coverage(),
summary.metahunt_conformal() — conformal
diagnostics.minmax_regret() — covariate-free
worst-case-regret baseline (Zhang, Huang, & Imai 2024).reconstruction_error_curve(),
cv_error_curve() — rank-selection
diagnostics.select_denoising_params() —
cross-validated tuning of (N, Delta).f_hat_from_models(),
build_grid() — onramp from fitted-model lists to
the package’s matrix inputs.The individual pipeline building blocks (dfspa(),
project_to_simplex(), fit_weight_model(),
predict_target(), apply_wrapper()) are also
exported and can be composed independently.
If you use MetaHunt in academic work, please cite:
Shi, W., Imai, K., and Zhang, Y. (2024). Privacy-preserving meta-analysis through low-rank basis hunting. arXiv:2604.23847. https://arxiv.org/abs/2604.23847
v0.1.0 is the first GitHub release. The API is
stabilising; user feedback is welcome via issues.
MIT © Wenqi Shi, Kosuke Imai, Yi Zhang
These binaries (installable software) and packages are in development.
They may not be fully stable and should be used with caution. We make no claims about them.