afcp(): thin Average Feature Choice Probability estimator over a matchup. se_type now defaults to NULL, deferring to lm_robust()’s own default (CR2 when clustered, HC2 otherwise); the previous hard-coded "stata" silently changed the standard errors. New min_clusters argument (default 2) drops groups with too few clusters to estimate, matching the study-level behavior in the meta-analysis pipeline. Validated to reproduce the paper’s Shiny-app estimates bit-for-bit on the full 163-study corpus.
afcp() returns n_clusters (distinct clusters, i.e. respondents) alongside n (observations, i.e. choice tasks). lm_robust() was already computing the cluster count on the fit being summarised and the tibble discarded it, which left a caller no way to report respondents at all.
Bug fix: afcp() now tests against a null of 0.5 rather than 0. The AFCP is a choice probability, so the coin-flip null is the one of interest; the intercept test lm_robust() reports is against 0, which made p.value essentially 0 for any well-powered study, including one whose estimate sat exactly at 0.5. New null_value argument (default 0.5) sets the null; pass null_value = 0 to recover the old intercept test. estimate, std.error and the confidence interval are unchanged.
Bug fix: get_matchups() with an unconstrained side (an empty A or B list) no longer errors with phantom _1/_2 columns (a paste0() zero-length recycling gotcha).
Bug fix: afcp() now drops a degenerate study (e.g. a single cluster) whose fit returns unusable components, instead of erroring during coefficient extraction; the tryCatch covers extraction, not just the model fit.
Test coverage raised to ~99% (added tests for weights, unconstrained-side matchups, degenerate-fit handling, and the input-validation error paths).