LoginSign Up
← Back to all sparks
P

plssem

INFRA · APIS
Velocity0.0

Complex Partial Least Squares Structural Equation Modeling

github.com/kss2k/plssem ↗

plssem took PLS-SEM into multilevel data, then spent two releases making the estimates trustworthy.

structural-equation-modelingpartial-least-squaresmultilevel-modelsstandard-errorsr-packages
◆Current state
plssem is a young R implementation of partial least squares structural equation modelling, three CRAN releases old and shipping monthly. Its distinguishing work is the MC-PLS family — consistent PLS estimators the maintainer extended to mixed-effects designs in June — and the releases since have been about getting standard errors, admissibility and fit measures onto the same footing as the point estimates.
◆Where it's heading
The pattern is capability first, inference second. Multilevel MC-PLSc and MC-OrdPLSc arrived in 0.1.2 together with Monte-Carlo delta-method standard errors and a Polyak-Juditsky extrapolation step; 0.1.3 then extended delta-method errors to redundant parameters and thresholds, optimized their computation, added a loglikelihood-based fit measure and generated dynamic bounds to keep MC-PLS solutions admissible. Admissibility recurs throughout — penalized inadmissible solutions in 0.1.1, variance lower bounds and negative residual variance handling in 0.1.3, and an option to drop inadmissible bootstraps rather than silently include them. The release notes are pull-request lists, so the reasoning behind each change stays in the repository.
◆Prediction
The MIMIC mode and GLS estimator both landed in the most recent release without the standard-error and fit-measure work that followed earlier additions, so extending inference to cover them is the natural next step. Bootstrap defaults moving to 500 replications suggests runtime is a live constraint and further optimization is likely.

◆Recent moves

  1. 1mo ago

    MIMIC mode, a GLS structural estimator and delta-method thresholds

    The inference catch-up release for the multilevel work: delta-method standard errors extended to redundant parameters and thresholds and then optimized, a loglikelihood-based fit measure for MC-PLS, and dynamic bounds that keep solutions admissible. MIMIC mode for constructs carrying both reflective and formative indicators and a GLS structural estimator widen specification without changing the package's direction.

    View source ↗
  2. 2mo ago

    MC-PLSc and MC-OrdPLSc extend to multilevel and mixed-effects models

    ⚡ SPARK

    The release that changes what data plssem can be pointed at. Everything the package could previously estimate assumed a single-level sample; MC-PLSc and MC-OrdPLSc gain mixed-effects forms, and the surrounding work — Monte-Carlo delta standard errors, Polyak-Juditsky extrapolation, a faster biased GLS estimator — exists to make those fits converge and report defensible uncertainty.

    View source ↗
  3. 3mo ago

    Parallel bootstrapping, kNN and mean imputation, higher-order constructs

    The build-out release before the multilevel work: parallel bootstrapping and optimized MC-OrdPLSc bootstraps address runtime, kNN and mean imputation address missing data, and higher-order constructs and prediction fill out the modelling surface. Penalizing inadmissible solutions starts a thread that runs through every later release.

    View source ↗