LCPA 1.0.4

• We gratefully thank Sungbo Sim (pposam@naver.com) for using LCPA and for the careful, detailed issue reports and suggestions that helped improve this release.
• Unified LCPA() and LTA() around type.analysis = "XZ" and "ZY". The measurement model is selected by type.model, its estimator by method.model, and its detailed settings by control.model.
• Added ML and BCH three-step analyses for covariate effects on latent groups (XZ) and latent-group effects on continuous or categorical dependent variables (ZY). Regression estimation and standard errors can be analytic, numerical, or bootstrap-based; continuous dependent variables must be standardized before analysis.
• Extended LTA with state-based and complete-path ZY analyses, time-specific or time-invariant effects, optional pooled Step 1 estimation, participant-level bootstrap, and compiled forward-backward evaluation.
• Expanded ZY results to report conditional means, variances, category probabilities, standard errors, confidence intervals, and omnibus Wald tests. Step 3 sandwich standard errors now account for Step 2 CEP estimation uncertainty where applicable, which can change standard errors relative to earlier releases.
• Reimplemented LRT.test.VLMR() using the robust-sandwich weighted chi-square reference distribution for Mplus TECH11, evaluated by Imhof’s method while retaining negative eigenvalue weights. The adjusted LMR uses the general parameter-difference correction from Lo, Mendell, and Rubin (2001).
• Improved the bootstrap likelihood-ratio test so refits inherit the original fitting controls, with sequential stopping and diagnostics for negative bootstrap LRT statistics.
• Corrected the LCA parameter count, two-sided p-values and confidence-interval rounding, CEP orientation and normalization, LCA EM synchronization, polytomous smoothing, and several likelihood, posterior-probability, model-selection, and simulation-label calculations.
• Improved analytic-gradient optimization, boundary and singular-matrix diagnostics, Louis-information and sandwich standard errors, and standardized fitted-object, simulation, progress, and S3 output. Optional SEM backends through flexmix, Rmixmod, and RMixtComp and the LCPA/LTA documentation and examples were also expanded.

LCPA 1.0.3

• Improved NNE performance and corrected the LPA EM algorithm.
• Corrected documentation errors.

LCPA 1.0.2

• Added adjust.model() for aligning LCA or LPA solutions and plot() methods for both model types.
• Improved Mplus variable handling and removed unnecessary Mplus statements.
• Improved Python dependency selection and GPU-based neural-network estimation.
• Improved bootstrap likelihood-ratio testing and corrected documentation errors.

LCPA 1.0.1

• Added use.attention control for neural-network estimation.
• Corrected control.NNE configuration and LCA parameter names.

LCPA 1.0.0

• Initial release