seminr

• In predict_pls(), a PLSc solution can be inadmissible in a training fold even when the full-sample solution is admissible, especially with few folds (for the mobi ECSI model, in about 70% of 5-fold and 30% of 10-fold runs). Such folds are skipped with one warning that names them: their test rows are NA for both PLS and the LM benchmark, and the prediction metrics use the remaining rows. If every fold is inadmissible, predict_pls() stops. The PLS chain is never used as a fallback. Cross-validation also no longer turns any error or warning inside a fold into a failure with an unrelated message ("row names supplied are of the wrong length"); they now reach the caller.
• Models estimated by seminr < 2.6.0 and saved (e.g. with saveRDS()) keep their old PLSc estimates. predict() on such a model that mixes reflective() constructs with Mode A or unit-weighted composites now warns to re-estimate it. estimate_pls() stamps models with seminr_version for this purpose.
• PLSc (estimate_pls() with reflective() constructs) corrected the paths into and out of Mode A and unit-weighted composites as if they were common factors, by dividing their correlations by sqrt(rho_A). Only Mode B composites were left uncorrected. Consistent PLS corrects only common factors and treats composites as fully reliable (Dijkstra & Henseler, 2015), as cSEM does. All composites now use rho_A = 1 in PLSc. **Path coefficients and R² change for models that mix reflective() constructs with Mode A or unit-weighted composite() constructs.** Loadings, all-reflective models, models with only Mode B composites next to reflective constructs, and models without reflective constructs are unchanged. The rho_A reported by summary() is unchanged. Calling PLSc() directly on a model without reflective constructs now leaves it unchanged (it used to disattenuate its composites) (#430).
• PLSc corrected a two-stage higher_composite() with rho_A computed on its lower-order construct (LOC) scores, as if the HOC were a common factor of them. A HOC of composite LOCs is now treated as fully reliable (rho = 1). A HOC of common-factor LOCs is corrected with the reliability of its stage-2 proxy, a weighted sum of error-laden LOC scores: w'S*w, with the LOCs' stage-1 reliabilities on the diagonal of S* (van Riel et al., 2017), as cSEM's two-stage approach does. **Path coefficients and R² change for PLSc models with higher_composite() constructs** (on mobi, IE -> Satisfaction 0.953 -> 0.920 with reflective LOCs; cSEM 0.918). A HOC of common factors is now corrected even when no stage-2 construct is reflective, and a reflective() construct that is declared but not used in the structural model no longer makes estimate_pls() run PLSc.
• predict_pls() on a model from estimate_cbsem() failed with an unrelated low-level error (the CB-SEM object inherits seminr_model but has no weights or loadings). It now stops with a message pointing to predict(). predict() and predict_pls() on a model from estimate_cfa() failed the same way; they now stop and say that a CFA has no structural model to predict from (#426).
• plot.reliability_table() drew its reference line at 0.708, which is the indicator **loading** threshold (0.708² ≈ 0.50 explained variance). The metrics that plot shows — Cronbach's alpha, rhoA and rhoC — are construct-level internal consistency reliabilities, judged against **0.70**. The line is now drawn at 0.70, and the value is exposed as a threshold argument so it can be overridden and so the intended value is documented rather than buried (#421). Reported by Marko Sarstedt.
• predict_pls(reps = k) repeated the cross-validation on the same folds, so every repetition gave the same predictions and reps had no effect. Each repetition now draws new random folds, and the predictions are averaged over the repetitions. **Results change for any call with reps greater than 1**, including assess_cvpat(), assess_cvpat_compare(), assess_pcm() and assess_coa() in seminrExtras, which pass reps on to predict_pls(). reps = NULL and reps = 1 give the same results as before. A fold skipped as inadmissible (PLSc) in one repetition is averaged over the others.
• predict() on test data with a missing value returned NA for every prediction in that case's row, because the matrix products spread the NA through zero weights, loadings and paths (NA * 0 is NA). Only the predictions that use the missing value are NA now: a missing item makes its own construct's score NA and the predictions that depend on it, and leaves the others. Predictions for complete data are unchanged. PLSc models already behaved this way.
• estimate_pls() returned inadmissible PLSc solutions without notice. It now warns, once, with every problem found: a rho_A above one, a standardized loading above one, a disattenuated construct correlation of 1 or more in absolute value, a structural-model-implied construct correlation matrix that is not positive definite, or an R-squared outside [0, 1]. The estimates are still returned. Such solutions are common in small samples.
• bootstrap_model() drops resamples whose estimation fails or warns, so inadmissible PLSc resamples are now excluded from the bootstrap statistics. The number dropped is reported in the final message, stored in the model as boots_dropped (with boots_requested), and printed by summary(). **Bootstrap standard errors, t values, p values and confidence intervals change for PLSc models with inadmissible resamples**: for a mobi model with a weak reflective block, 69 of 500 resamples were dropped, and the t value of its weakest discriminant validity estimate rose from 0.95 to 1.52.
• PLSc divides by sqrt(rho_A). A rho_A of zero or below gave NaN paths, R-squared and loadings with only R's "NaNs produced" warnings; estimate_pls() now stops and names the constructs. A cross-validation fold or bootstrap resample with such a rho_A is skipped.
• construct_order() looped forever on structural models in which an outcome construct's antecedents are ordered before the rest of the model (e.g. Image -> Loyalty next to Image -> Expectation -> Quality), so predict(technique = predict_EA) hung on them. A structural model with a cycle now stops with an error.
• The note that PLSc estimates of interaction models are subject to some bias (Becker et al., 2018) never showed, because estimate_pls() checked a field that is never set. It now shows for PLSc models with interaction terms.
• estimate_pls() with a higher_reflective() construct, which is supported by estimate_cbsem() only, failed with "undefined columns selected". It now stops and points to estimate_cbsem() or higher_composite().
• PLSc() on a model with a higher-order construct, already corrected by estimate_pls(), failed with "subscript out of bounds". It now returns the model unchanged with a message.
• Parallel processing failed to start ("creation of server socket failed") when its port was in use, e.g. by another R session started at the same time. It now retries on other ports.

Changed

• The package maintainer address is now seminrgroup@gmail.com (#420).
• bootstrap_model() with the default cores = NULL still uses every core. During R CMD check with CRAN's core limit set (_R_CHECK_LIMIT_CORES_, as with --as-cran and on CRAN), it uses at most two, so the package's tests and vignettes stay within CRAN policy. Results do not depend on the number of cores.

Performance

Substantial internal speedups across the heavy routines. **These are refactors

only — no result changes.** Verified against the PLS-SEM R book code: with seeds

pinned, 2.5.0 and 2.6.0 agree to max|diff| = 0 on every deterministic quantity

(path coefficients, loadings, weights, R², reliability, HTMT, AVE, VIF) and every

bootstrap quantity, for both cores = 1 and cores = 2.

• simplePLS() skips redundant matrix work in the estimation loop.
• HTMT() reuses one correlation block per construct instead of recomputing.
• predict_pls() avoids redundant model refits and per-item lm fits.
• Cross-validation reuses the full-model reference scores.
• PLS-MGA p-values are computed without expand.grid.
• bootstrap_model(cores = 1) now runs replicates in-process rather than shipping them to a one-worker PSOCK cluster, and saves and restores the caller's RNG state. Replicates continue to use set.seed(seed + i), so bootstrap results remain identical across core counts and across versions.
• Added a benchmark harness covering the heavy PLS routines.