Changes in version 0.3.0 (2026-10-01)                  

  - cglm() and cgam() gain a direction argument for the stepwise search,
    "forward" (the default and the previous behaviour) or "backward".
    The backward search starts from the full model and removes, at each
    step, the variable whose removal gives the largest Pearson-risk
    p-value, stopping once the model is no longer rejected. Both
    directions then prune by BIC as before. Forward remains the default
    because it is cheaper.
  - For a binomial response whose candidate set contains a categorical
    variable, the stepwise search now runs backward and says so. The
    Pearson risk of a binary regression on categorical covariates alone
    is mathematically equal to 1, so a forward search can enter such a
    variable at the first step and stop there.
  - The stepwise search now emits the same "only one categorical
    variable" message that the exhaustive search already emitted.
  - The backward search reports "no potential causal model found" when
    no model on its path was accepted, matching search = "all".
    Previously it returned the single-term model it happened to end on,
    which the test had rejected.

                        Changes in version 0.2.2                        

  - Parallel model search (ncores > 1) now works on all platforms. On
    Unix/macOS it continues to use forking (parallel::mclapply); on
    Windows it uses a PSOCK cluster (parallel::parLapply) with
    parallel-safe RNG streams. Previously ncores was silently ignored on
    Windows.
  - Added the causalreg.parallel option to select the parallel backend
    ("auto", "fork", "psock", "sequential").

                        Changes in version 0.2.1                        

  - cgam() fast bootstrap via fixed smoothing parameters (fast_gam =
    TRUE, the default): the smoothing parameters selected on the
    original data are held fixed across that model's bootstrap
    resamples, giving a 3-4x speedup.

                        Changes in version 0.2.0                        

  - Added C++/Rcpp acceleration for GLM fitting, Pearson statistics, and
    the bootstrap (use_cpp = TRUE, the default for supported families).
  - Added parallel model evaluation via the ncores argument.