Predicting Continuance Intention in YouTube-Based Informal Learning: A Random Forest Regression and Linear Analysis

Structural associations and necessary conditions for continuance intention do not, by themselves, establish whether a nonlinear-capable model improves prediction over a linear benchmark or whether predictor rankings remain stable across data partitions. This secondary analysis addresses these questions in YouTube-based informal learning using existing data from 797 Thai learners. PLS supported measurement assessment and descriptive linear estimation, while predictive models used fixed, unweighted item means. Random Forest Regression (RFR) and ordinary least squares (OLS) were assessed on matched 80/20 partitions, 10 × 10-fold cross-validation, and 100 repeated 80/20 splits. Original-test R2 was 0.6625 for RFR and 0.7036 for OLS; RFR RMSE and MAE were 0.5839 and 0.4050. Across 100 splits, the mean R2 was 0.6382 for RFR and 0.6549 for OLS (paired difference = −0.0168; corrected p = 0.1048), providing no demonstrated predictive advantage for the evaluated RFR configuration. In 30 importance-sensitivity partitions, Satisfaction ranked first throughout, while Perceived Enjoyment and Trust recurrently occupied the top three. Conditional-permutation diagnostics limited claims of unique contributions. The nonlinear-capable RFR analysis complements linear estimates by showing Satisfaction’s consistently highest marginal predictive importance across the assessed partitions.