Why engagement fails at the classroom level: multilevel evidence from AI-supported learning in higher education

Description This dataset accompanies the study “Why engagement fails at the classroom level: Multilevel evidence from AI-supported learning in higher education.” It provides a comprehensive multilevel dataset designed to examine how student-level cognitive processes and classroom-level instructional context jointly shape learning outcomes in artificial intelligence-supported environments. The dataset was collected from undergraduate students across six universities within Vietnam National University, Hanoi (VNU Hanoi) during the 2025 academic year. Data collection was conducted from March to June 2025, corresponding to the latter half of the semester, ensuring that students had accumulated sufficient experience with AI-supported learning activities before reporting their perceptions and behaviors. The dataset follows a hierarchical structure in which students are nested within intact classes, while classes are situated within universities used as sampling strata, enabling robust multilevel analysis across both instructional and institutional contexts. Data Structure The dataset is organized in a wide-format structure, where each row represents an individual student and each column represents a variable measured at the student level or derived from higher-level contextual structures. The hierarchical nature of the dataset is captured through both class identifiers and university identifiers, allowing for flexible modeling of nested data structures. The dataset consists of five main components: First, identifiers and background variables define the hierarchical structure and control variables. These include a unique anonymized student ID, a university identifier representing six institutions within VNU Hanoi, and a class identifier reflecting the nested structure of students within classrooms. Prior academic performance is measured using GPA on a 10-point scale and serves as a baseline control variable in subsequent analyses. Second, artificial intelligence self-efficacy (AISE) is measured at the student level using four Likert-scale items (1–7). These items capture distinct dimensions of AI-related competence, including operational confidence in using AI tools, evaluative confidence in assessing AI-generated outputs, integrative confidence in embedding AI into learning workflows, and adaptive confidence in adjusting prompts and strategies. A composite score is calculated as the mean of the four items, representing overall AI self-efficacy. Third, cognitive offloading (COFF) is measured using four Likert-scale items (1–7), capturing the extent to which students delegate cognitive processes to AI systems. The items represent multiple domains of delegation, including memory retrieval, idea generation, problem-solving, and writing support. A composite score is provided to reflect overall reliance on cognitive offloading in AI-supported learning contexts. Fourth, student engagement (ENG) is measured using four Likert-scale items (1–7), reflecting multidimensional involvement in learning activities. These items capture behavioral engagement, emotional engagement, cognitive engagement, and agentic engagement, ensuring a comprehensive representation of learning behavior. A composite score represents the overall level of engagement. Fifth, teacher support (TSUP) is measured using four Likert-scale items (1–7) based on student perceptions of instructional practices. These items capture guidance, responsiveness, encouragement, and availability of instructors. Although measured at the student level, this construct can be aggregated to the class level to represent shared instructional context, enabling cross-level analysis. Finally, the dataset includes an academic outcome variable, semester GPA, measured on a 10-point scale. This variable allows examination of how cognitive processes and engagement translate into academic performance across both individual and classroom levels. Measurement Characteristics All psychological constructs are measured using 7-point Likert scales, allowing fine-grained assessment of students’ perceptions, behaviors, and cognitive strategies in AI-supported learning environments. Each construct is operationalized using four reflective indicators grounded in theoretically distinct subdimensions, ensuring strong psychometric validity and construct differentiation. Academic performance variables are measured using a 10-point GPA scale, consistent with grading practices in Vietnamese higher education. Composite variables are computed as mean scores for descriptive purposes, while item-level indicators are used for latent variable modeling in structural equation modeling frameworks. Missing Data The dataset contains minimal missing data due to structured classroom-based data collection procedures. Where missing values occur, they are coded as 999 to ensure compatibility with Mplus and other SEM software. This coding scheme can be specified using: MISSING = 999; Researchers are encouraged to apply full information maximum likelihood or equivalent robust estimation methods when handling missing data in multilevel analyses. Analytical Scope The dataset supports a wide range of advanced statistical analyses, particularly those involving hierarchical data structures. It is especially suitable for multilevel structural equation modeling, allowing simultaneous estimation of within-level (student) and between-level (classroom) relationships, while also accommodating institutional variation across universities. At the student level, the dataset enables examination of cognitive–behavioral mechanisms linking AI self-efficacy, cognitive offloading, engagement, and academic performance. At higher levels, the dataset allows investigation of how instructional context and institutional variation influence collective engagement and academic outcomes. In addition, the dataset supports mediation analysis, indirect effect estimation, and cross-level modeling, providing a robust framework for understanding how individual cognition and contextual conditions interact in AI-supported learning environments. Reproducibility The repository includes a detailed codebook, Mplus syntax files, and corresponding output files for all major analyses. These materials enable full replication of the analytical procedures reported in the study, ensuring transparency and reproducibility. Contribution This dataset contributes to the growing body of research on artificial intelligence in education by providing multilevel empirical evidence from a large-scale higher education system in Vietnam. The inclusion of multiple universities within a unified institutional system enhances generalizability while maintaining consistency in curricular and instructional conditions. The dataset is particularly valuable for researchers interested in AI-supported learning, cognitive offloading, student engagement, and multilevel modeling approaches. It offers a rigorous empirical foundation for examining how learning processes are shaped by the interaction between individual cognition, classroom dynamics, and institutional context.