NEWS

lpmec 1.3.0

Split-based corrected OLS now uses score-scale (Spearman-Brown) reliabilities. In split_scores and observables modes the regression uses the full score (the average of the two z-scored halves, or the full-battery estimate) while the raw half-split correlation estimates the reliability of a half score, so the previous correction b * sqrt(r_pooled) / r_design overcorrected (by tens of percent under aggressive design transforms). The split variant now applies b * sqrt(SB(r_pooled)) / SB(r_design) with SB(r) = 2r / (1 + r), matching Assumption 3 of the accompanying paper and agreeing asymptotically with the corrected split-IV. New fields split_rho_score / design_split_rho_score report the stepped-up values; the raw correlations are still returned. The triad, pair, and split-IV corrections were already correctly scaled and are unchanged, as is the two-measure scores-mode path (the V2 reference pipeline).
Cross-measure IV corrections are now target-oriented (Proposition 3c). With 3+ measures, each m_by_l coefficient is multiplied by sqrt of target m's pooled triad reliability instead of sqrt of the pairwise correlation r_ml (consistent only under equal reliabilities). The two-measure parallel-pair fallback is unchanged.

Latent outcomes (Proposition 4). lpmec_panel_onerun() and lpmec_panel() accept Y_split_scores (an n x 2 matrix of outcome half scores): the outcome reliability rho_Y is estimated by Spearman-Brown step-up and every corrected estimator is additionally divided by sqrt(rho_Y). Y becomes optional (the outcome score is then built from the halves), and unit may be NULL for design = "pooled", covering the pure cross-sectional case.
Design-local scale (Proposition 2a). The panel functions report sd_design_x, the naive local slope ols_coef_local (effect per SD of the design-transformed latent trait), and corrected local coefficients corrected_ols_coef_local(_split/_triad/_pair) with their sensitivity range, satisfying the exact local-pooled bridge identity.

Y_split_scores is inserted after covariates in the signatures of lpmec_panel_onerun() and lpmec_panel(); calls that passed design or later arguments positionally must switch to named arguments.

lpmec_reliability_bounds() and the panel reliability tables gain a rho_split column (Spearman-Brown step-up of the split correlation to the full-score scale); rho_lo / rho_hi now span {triad, rho_split} so both candidates estimate the same full-score target.
Documentation reframed per the paper's Proposition 7: the [rho_lo, rho_hi] interval is a sensitivity range, not a partial-identification interval, unless the corresponding orthogonality or ratio condition is maintained; directional claims about the split and triad candidates are now stated with their conditions.
Proposition numbering in the documentation updated to the current manuscript (design correction = Prop 3, latent controls = Prop 5, latent moderators = Prop 6, triangulation = Prop 7).

Added lpmec_panel_onerun() and lpmec_panel() for measurement error correction under panel designs ("within", "twoway", "difference", "pooled"): NA-safe design transforms (iterated demeaning for unbalanced two-way panels; exact-gap k-period differencing), cluster-robust naive OLS, within- and cross-measure split-IV, and the design-aware correction algebra (corrected OLS is b * sqrt(rho_pooled) / rho_design; corrected split-IV multiplies by sqrt(rho_pooled) -- IV multiplies, OLS divides).
lpmec_panel() reports cluster (unit) bootstrap uncertainty with fresh pseudo-ids per draw, re-running the whole pipeline (scoring, sign alignment, transforms, reliabilities, regressions, corrections) on every (bootstrap, partition) pair.

Added lpmec_reliability_bounds(): per (measure, design) split correlations and cross-measure triad reliabilities (rho*_m = r_ml * r_mk / r_lk, requiring 3+ measures), with the implied identification interval [rho_lo, rho_hi] and an optional cluster bootstrap. Reliabilities below min_reliability and correlations on fewer than min_cor_n complete pairs are reported as NA rather than divided out.

Added lpmec_moderator_onerun() and lpmec_moderator() for treatment-by-latent-moderator interactions in experiments: the naive interaction coefficient is divided by the square root of the Spearman-Brown as-used score reliability (rho_score = M * rbar / (1 + (M - 1) * rbar)), with per-measure triad corrections when 3+ measures are supplied and a row (or stratified) bootstrap that propagates reliability-estimation uncertainty into the corrected-interaction intervals.

The new functions share a tri-source measure-input model merged by name: observables (raw item batteries scored via lpmec_onerun()), split_scores (n x 2 half-score matrices, required for measures with fewer than 4 items), and scores (pre-computed full scores). All scores are pooled z-scored and sign-aligned before any correlations are taken.

Added print()/summary() methods for lpmec_panel_onerun, lpmec_panel, lpmec_moderator_onerun, and lpmec_moderator, print.lpmec_reliability_bounds(), and plot() methods for lpmec_panel and lpmec_moderator.

Vignette gains three sections: panel/fixed-effects corrections, reliability bounds with multiple measures, and latent moderators in experiments.

Fixed .lpmec_aggregate_by_boot() dropping the matrix structure for single-column inputs, which broke bootstrap aggregation for lpmec_multivariate() runs with a single latent predictor.

Added m-out-of-n and subsampling uncertainty for nonsmooth finite-partition median aggregation. The implementation fixes the realized partition set across resamples, reruns the full latent measurement and correction pipeline within each resample, and reports root-scaled standard errors and confidence intervals for m < n designs.

Renamed package from lpme to lpmec (Latent Predictor Measurement Error Correction).
The name "lpme" was already taken by a different archived CRAN package (by Zhou & Huang).
Main functions are lpmec() and lpmec_onerun().

Fixed vignette to use correct estimation_method values ("em" instead of "emIRT").
Added .Rbuildignore to exclude development files from package build.
Replaced eval(parse()) pattern with direct assignment for CRAN compliance.
Changed F/T to FALSE/TRUE throughout codebase.
Wrapped examples in \donttest{} to avoid CRAN timeout issues.
Changed conda_env_required default to FALSE for CRAN compatibility.
Updated CITATION file to use modern bibentry() format.
Added skip_on_cran() guards to test files.
Added internal function documentation with @noRd tags.

Fixed documentation to correctly indicate pscl as the default MCMC backend.
Refreshed public documentation for advanced estimation methods, MCMC controls, return-value fields, and current arXiv links.
Updated version year in CITATION file.