ccva 0.1.1

New features

• ccva_combine(method = "rule") implements the rule-based framework of Foden et al. (2013): indicators flag units as high exposure, high sensitivity or low adaptive capacity (coded 0/1 indicators at their risk end, continuous ones by a threshold quantile), a component is flagged if any indicator is, and units get Foden’s categories (highly vulnerable, high latent risk, potential adapter, potential persister). Missing indicators are treated as unknown = "optimistic" or "pessimistic". details$drivers and ccva_plot(fit, "effects") show how many units each indicator flags on its own.
• Missing values: as_ccva(na_action = "keep") keeps incomplete units. The "mean" and "ordinal" scorers average the observed indicators, and the latent model uses full-information maximum likelihood, so units with a missing outcome (e.g. IUCN Data Deficient species) are scored from their indicators. Diagnostics, validation and calibration use the available data; summary() and print() report missingness.
• Formative components with estimated weights: ccva_combine(method = "latent", formative_weights = "estimated") enters every indicator of a formative component as a cause of latent vulnerability (a MIMIC model) with its own sign-constrained weight, reported in details$indicator_weights and plotted by ccva_plot(fit, "effects").
• Ordered categorical outcomes: ccva_combine(method = "latent", ordered =) fits outcomes such as IUCN categories as ordinal (WLSMV).
• ccva_cv() computes cross-validated scores: each fold is scored by a model refitted without it, with the outcomes of the held-out units hidden (as for an unassessed species). With cluster, whole groups (families, basins) are kept in one fold. ccva_calibrate(cv = ) uses these scores for out-of-sample calibrated risk, and returns probability draws for intervals on sums such as the expected number of threatened species.
• Cluster bootstrap: ccva_uncertainty(cluster = ) resamples whole groups of related or co-occurring units instead of single units, so intervals respect dependence within groups.
• Bootstrap and cross-validation refits of the latent model skip fit measures, which makes them several times faster with missing data.

Bug fixes

• Shape-constrained fits now compute vulnerability scores with predict() on the fitted model rather than from the fitted values stored by scam, so that scores and predict() on the same data agree on every platform (fixes a test failure on r-oldrel-macos-arm64).

ccva 0.1.0

First release.

• Data contract as_ccva(): any dataset, columns tagged to exposure, sensitivity and adaptive capacity, with directions and outcome proxies.
• Simulator ccva_simulate() with a known latent structure (linear or nonlinear truth, optional genomic block) and bundled dataset ccva_sim.
• Component scoring: Ofori-style ordinal, continuous mean, PCA and sign-constrained penalised regression.
• Combination models: weighted index (Ofori et al. 2017, corrected normalisation), shape-constrained additive model (scam), constrained latent SEM (lavaan) with shared-error method factors, and penalised two-stage model (glmnet).
• Constraint system on predictors and components: sign, bounds, equality, fixed values and adaptive capacity as moderator.
• Diagnostics, validation, method comparison, bootstrap and weight uncertainty (with factor-score error for latent models), calibrated classification, plots for every step and an HTML report.
• Method registry ccva_register() for contributed methods.
• Framework tools: directed links between components (constrain_link()) with direct, indirect and total effects, or two-stage resolution; formative components (as_ccva(formative = )) and mean inter-item correlation in ccva_check(); concave / convex shapes (constrain_shape()); calibrated risk predictions (ccva_calibrate()); method recommendations by data availability (ccva_recommend()); assay methodology adjustment (ccva_adjust_method()); community / site aggregation (ccva_aggregate()); workflow vignette “framework”.