PTSDdiag

• bootstrap_stability() re-runs the entire exhaustive search on bootstrap resamples of the respondents and reports, for every combination, how often it stayed within delta of that replicate's best (pi) and how often it was the outright winner (p_argmax, with ties shared equally). Taking the reference point within each replicate rather than from the original sample keeps every comparison on the scale of its own resample. Note that pi for the combination that won in the original sample is optimistic, since replicates share most of their observations with the sample that selected it; the documentation says so. Per-replicate plateau sizes are returned as well, for a percentile interval across replicates.
• symptom_selection() reports how often each item appears across the plateau against the chance baseline for that rule form, computed by enumeration rather than assumed. Raw selection counts are zero-sum and cannot be read alone. The baseline is flat at n_symptoms / 20 without a cluster constraint, but under the hierarchical rule it ranges from 25.2% for each cognition-and-mood item to 55.8% for each avoidance item, so a flat comparison would make two-item clusters look preferred when they are merely unavoidable. A stability weighted variant is available when a bootstrap_stability() result is supplied.
• plot_symptom_selection() draws those frequencies with the chance baseline marked item by item, or as a ratio to chance.
• compare_rule_forms() runs the above across a named list of searches and stacks the results into long data frames with a rule_form column, ready for a manuscript table.

Multi-site transport and release

• evaluate_sets() scores a supplied list of symptom combinations against the full DSM-5-TR PCL-5 diagnosis. Where score_all_combinations() enumerates and scores a whole candidate space, this scores exactly the combinations it is given, which is what a validation site needs when the search happened elsewhere. It uses the same collapsed-pattern matrix path as bootstrap_stability(), so cost scales with the number of distinct response patterns rather than with sample size, and it reproduces score_all_combinations() exactly.
• transport_plateau() turns those per-combination results into something a site may release: a five-number summary of balanced accuracy, sensitivity and specificity **across** each transported plateau, plus two-by-two counts for a short, pre-specified list of named combinations. The combination-to-value linkage is discarded deliberately — per-combination counts across a whole candidate space are enough constraints to recover the response-pattern distribution, and in a small sample a rare pattern is effectively an individual record. The function has no argument that emits per-combination detail for a plateau, so it cannot be released by accident.
• write_transport() and read_transport() serialise a return bundle. Every filename is prefixed with the sample identifier and every table carries sample_id and payload_id columns, so bundles from different sites stay distinguishable as loose files in one directory. dir is required and must already exist, matching write_combinations(): the package never picks a location for you and never creates directories, so a mistyped path fails instead of scattering files. session_info is off by default.
• diagnostic_metrics() computes every metric a two-by-two table supports, with intervals, recomputed from the counts. The interval functions behind it are exported too: wilson_ci(), exact_ci() (Clopper-Pearson, for small denominators), ba_ci(), lr_ci() (Simel) and kappa_ci(). ba_ci() combines the component intervals by Newcombe square-and-add (see Bug fixes), and every interval here is documented as **marginal**: none is a valid basis for comparing two combinations, which are paired on the same individuals.
• five_number() reduces a vector to the six numbers a site may release.
• icd11_items() exposes the six PCL-5 items create_icd11_diagnosis() scores, so a table describing the operationalisation is generated from the same definition that is applied and cannot drift from it.
• chance_baseline() and n_candidates() expose the size of a candidate space and the proportion of it containing each item. Under a cluster constraint the baseline is a step function of the cluster rather than a constant, because a two-item cluster contributes far more of its combinations than a seven-item one — which is what makes a raw selection frequency misleading on its own.
• set_id_to_items(), format_set_items(), format_set_labels() and pcl5_item_labels() convert between combination ids and item numbers or labels. as_set_matrix() accepts a combination specification in any of the three shapes the package produces.

Bug fixes

• ba_ci() no longer returns a zero-width interval when sensitivity or specificity is at a boundary. The analytic variance Var(BA) = 0.25 (Var(sens) + Var(spec)) is degenerate there, because p(1-p)/n is zero, so a table with six positives all classified correctly reported 1.00 (1.00, 1.00) — no uncertainty at all from 183 observations, while its own component intervals were correctly wide. The default is now Newcombe square-and-add of the component intervals, which is not degenerate at a boundary: the same table gives 1.00 (0.80, 1.00) under the default method = "wilson", and 1.00 (0.77, 1.00) under method = "exact". The analytic form stays available as method = "wald" and warns when a component sits at 0 or 1.
• diagnostic_metrics(ci = ) now reaches balanced accuracy. It previously applied the requested method to the proportions but always used the analytic form for their mean, so ci = "exact" returned Clopper-Pearson bounds on sensitivity and specificity next to a Wald interval on balanced accuracy. ba_method overrides.
• **The hierarchical rule could accept a set spanning one cluster too few.** When clusters did not cover all 20 items, .diagnose_combination() counted items belonging to no cluster as though they formed a cluster of their own: the lookup table was sized to the highest clustered item, so an endorsed item above that index read as NA and one inside the range but outside every cluster read as 0, and unique() then counted either as a distinct cluster. With clusters = list(B = 1:5, C = 6:7, D = 8:14), a respondent endorsing items 1, 6 and 15 was diagnosed under a 3-of-3 hierarchical rule despite spanning only two of the three clusters. Items outside every cluster now contribute nothing towards spanning. **Results are unchanged for the standard DSM-5-TR PCL-5 clusters**, which cover all 20 items, so no analysis using the defaults is affected. Affects optimize_combinations_clusters(), score_all_combinations() and apply_symptom_combinations() when custom, non-exhaustive clusters were supplied.

Improvements

• symptom_frequency() gains Baseline and Enrichment columns and a clusters argument, for the same reason: selection frequencies over a fixed number of combinations are only interpretable against chance. Existing columns are unchanged.
• plot_symptom_frequency() gains type = "enrichment", filling each tile with the ratio of observed to chance selection on a scale diverging around 1. The default remains type = "relative".

Notes

• The bootstrap collapses respondents to unique binarized response patterns and expresses each replicate as two matrix products, which is exact rather than an approximation because the reference diagnosis and every decision rule depend on the data only through the binarized responses. A full 2,000-replicate run over all 38,760 six-item combinations takes seconds.
• Under the hierarchical rule, items belonging to no cluster count towards the endorsement threshold but not towards spanning the clusters. This follows from the bug fix above and holds identically in the per-row rule and the matrix path, which are checked against each other on every run.