Hybrid PLS and machine-learning comparisons require explicit estimands and validation of the complete prediction pipeline. This methodological simulation evaluates a privacy-aware educational early-warning-system test bed in 40 distinct conditions, with 300 independently seeded replications each (12,000 datasets). Sample size ranges from 200 to 1,600, and nonlinearity varies across three functions and four levels, with the linear baseline counted once. Population scaling and standardized best linear projection targets replace unadjusted coefficient comparisons. PLS Mode A and consistent PLS (PLSc) are evaluated for parameter recovery. Prediction of simulated latent behavioral intention compares linear regression, tuned second-degree polynomial regression, a multilayer perceptron, and histogram gradient boosting. PLS weights and scaling parameters are fitted separately within every outer and inner training partition. A latent-information oracle is reported as a separate benchmark. Polynomial regression has the highest overall mean learner R^2 (0.573). Under quadratic nonlinearity at n = 1,600 and lambda = 0.60, gradient boosting yields Delta R^2 = 0.195 (95% CI [0.191, 0.199]) and Delta RMSE = -0.142 (95% CI [-0.145, -0.138]) relative to linear regression. PLSc reduces mean absolute path bias at n = 1,600 from 0.036 to 0.001. Practical gains are classified using a 0.02 $R^2$ threshold and familywise intervals; machine-learning superiority is not assumed. The results distinguish benefits from modeling curvature from benefits attributable specifically to complex learners, while separating simulation evidence from substantive validation.
Copyrights © 2026