                 Changes in version 0.3.0 (2026-08-25)                  

  - Added gls_mult() for feasible generalized least squares estimation
    under multiplicative heteroskedasticity. The estimator argument
    chooses the fit, either "ml" (default, Gaussian maximum likelihood)
    or "two_step" (Harvey's corrected auxiliary regression), and for
    maximum likelihood the method argument selects the stats::optim()
    algorithm: "BFGS" (default), "Nelder-Mead", "CG", or "L-BFGS-B". New
    coef() and vcov() methods access the mean and dispersion
    coefficients via model =, and confint(), tests(), summary(),
    print(), fitted(), and residuals() methods support applied
    inference, while logLik() and nobs() enable AIC() and BIC() for
    maximum likelihood fits. The package Description now also covers
    feasible generalized least squares following Harvey (1976) and
    Cribari-Neto and Pereira (2019).
  - Corrected the HC5 adjustment factor, which now follows the erratum
    to Cribari-Neto, Souza and Vasconcellos (2007). vcov_hc() and
    hcinfer() compute g_t = (1 - h_t)^(-delta_t / 2) instead of the
    expression g_t = (1 - h_t)^(-delta_t) printed in Section 3 of the
    original article, keeping delta_t = min(h_t / h_bar, max(4, k *
    h_max / h_bar)) and the default k = 0.7 unchanged. HC5 adjustment
    factors are therefore smaller, and HC5 standard errors, Wald
    statistics, p-values, and confidence intervals differ from those
    returned by earlier versions. HC5m is unaffected because it follows
    Li, Zhang, Zhang and Wang (2016) and applies its own exponent
    without the factor 1/2, so type = "hc5m" with k1 = 0, k2 = 0, and k3
    = 1 no longer reproduces type = "hc5" and instead squares its
    adjustment factor.
  - Documented the corrected HC5 expression in the vcov_hc() help page
    and in the methodology vignette, and added the 2008 erratum to the
    references of hcinfer(), vcov_hc(), tests(), the methodology
    vignette, and the package Description.

                 Changes in version 0.2.0 (2026-08-04)                  

  - Added boot_pairs() for pairs (case) bootstrap standard errors and
    confidence intervals of ordinary least squares coefficients. It
    resamples the observations with replacement, refits the model on
    each replicate, and summarizes the sampling distribution of the
    coefficients, providing an assumption-free empirical reference for
    the analytic heteroskedasticity-consistent standard errors from
    hcinfer() and vcov_hc(). Percentile, basic, and normal intervals are
    available, the resampling is reproducible through the seed argument,
    and the replicate fits can optionally run in parallel via
    purrr::in_parallel() and mirai without changing the numeric result.
  - Added coef(), vcov(), confint(), print(), and plot() methods for the
    hcinfer_boot objects returned by boot_pairs(). vcov() returns the
    bootstrap covariance matrix of the coefficients, confint() can
    recompute intervals at a different level or type directly from the
    stored replicates, and plot() draws the bootstrap confidence
    intervals, coloring each coefficient by whether its interval
    excludes or includes zero.
  - hcinfer() and vcov_hc() now accept independent HCbeta shape caps
    from 50 through 25000 inclusive, with defaults of 10000. HC0, HC1,
    and HCbeta also remain defined for an exact leverage value of one,
    while HC2, HC3, HC4, HC4m, HC5, and HC5m retain the positive
    leverage-complement requirement.
  - hcinfer() and vcov_hc() now enforce the fixed HCbeta shape floor
    of 0.01 after shrinkage and before the upper caps, including for
    nondefault leverage-complement truncation limits. The shape floor
    remains fixed when lower changes and is not a method argument.

                 Changes in version 0.1.1 (2026-07-13)                  

  - Added the PublicSchools2 dataset with 2024 per capita income, 2025
    public school expenditure per student, a Southern-region indicator,
    and complete variable and source documentation.
  - Standardized the federal district name in PublicSchools from
    Washington DC to District of Columbia.

                 Changes in version 0.1.0 (2026-06-10)                  

                     Changes in version 0.0.0.9000                      

  - Added the initial development version with HC covariance estimators,
    normal Wald inference, S3 output, and the PublicSchools dataset.
  - plot() now supports vcov_hc() objects, producing
    leverage-versus-adjustment-factor graphics for inspecting the
    relationship between h_t and g_t.
  - summary() now prints formal test results, confidence interval
    checks, and optional emoji markers to improve interpretation of
    robust inference output.
  - summary() now keeps displayed test_result decisions consistent with
    numeric p-values when p-values are displayed as <0.001.
  - Added tests() as a formal extractor for coefficient-level Wald test
    results. The function mirrors the API of confint(): an optional parm
    argument selects coefficients by name or position, and an optional
    alpha argument recomputes the reject column without affecting the
    stored p-values or test statistics.
