GitHub

Hypothesis Testing for Dependent Variables with Unbalanced Data.

HTDV provides a unified R toolkit for inference on dependent, unbalanced data under strong-mixing conditions, combining hierarchical Bayesian estimation via Hamiltonian Monte Carlo with frequentist and distribution-free robustness anchors (fixed-b HAR, block bootstrap, adaptive conformal).

Validation card

The framework is shipped with two pre-registered validation studies, both reproducible end-to-end and with their summary tables exposed as package datasets. See vignette("HTDV-validation").

  • Factorial Monte Carlo (htdv_sim_summary). 1024-cell design crossing sample size, AR(1) coefficient, innovation tail, imbalance ratio and location shift; 500 replications per cell × 3 inferential layers; 31 hours of wall-clock on 16 cores. The Bayesian envelope holds nominal size (mean 0.0556, sd 0.013) and nominal coverage (mean 0.944) across the entire grid; HAR and bootstrap inflate to empirical size 0.60 and coverage 0.29 in the worst corners under strong persistence. The asymptotic gap that motivates the framework is visible in the data.

  • External benchmarks (htdv_empirical_benchmarks). Three public datasets compared against published references:

    • FRED-MD post-1984 CPI inflation against Stock and Watson (2007).
    • Shiller log-CAPE against Campbell and Shiller (1998).
    • US-Canada 10-year yield differential against the iid Welch baseline.

    All three layers reproduce all three references with agreement in every case. The 95% interval widths scale monotonically with the series persistence: at

Read the original on github.com ↗