RESEARCH DOSSIER
Learning Analytics, Algorithmic Educational Decisions, and Student Data Governance
Abstract. Learning analytics organizes the collection, interpretation, and communication of educational data to support learning and teaching. This fundamental research monograph examines the psychological validity of educational trace data, predictive modeling and calibration, algorithmic decisions, student-risk interventions, fairness, causal effects, privacy, and student data governance. A critical narrative synthesis integrates foundational psychometrics, original educational and computational studies, systematic reviews, the 2026 learning analytics dashboard meta-analysis, and official standards. The evidence shows that predictive models can detect context-specific patterns, while educational benefits depend on decision policies and validated interventions. Average dashboard effects are modest and heterogeneous; original studies reveal unequal distribution of support under some predictive policies. The research develops evidence requirements for construct validity, external validation, prospective causal analysis, human review, contestability, and proportionate data governance, including human–artificial educational settings. The work does not report new experimental data or a protocol-registered systematic review.
Keywords: learning analytics; educational data mining; algorithmic educational decisions; educational psychology; student data governance; learning analytics dashboards; student success prediction; predictive calibration; educational risk classification; algorithmic fairness; student privacy; causal evaluation; self-regulated learning; human oversight; human–artificial education; Artificial Era
The Scientific Object of Learning Analytics and Educational Decision Governance
Learning analytics is the organized interpretation of evidence about learners, learning processes, and educational environments for decisions intended to improve learning and teaching. The SoLAR (2026) defines the field through collection, analysis, interpretation, and communication of data that yields theoretically relevant, actionable insights. The formulation links the computational activity to learning theory, human interpretation, and educational action. Its emphasis on communication and action matters because a mathematically accurate prediction that cannot guide a legitimate educational response remains incomplete as learning analytics.
The present research object comprises three interacting systems. The first is a measurement system converting observations of student activity into variables interpreted as achievement, participation, engagement, self-regulation, or risk. The second is an algorithmic decision system that converts measurements into forecasts, classifications, recommendations, resource allocations, or individual interventions. The third is a governance system defining purposes, authorized actors, accountability, rights, data provenance, access, retention, error correction, and review. Scientific analysis must trace the full pathway from a classroom event to an administrative consequence. Each link can introduce information loss, interpretive ambiguity, selection effects, or a new source of inequality.
A student opening a digital chapter late at night produces an event timestamp. The record establishes an interaction with a platform at a time; it does not independently establish whether the student understood the chapter, was distracted, had access to a quiet study environment, or was learning from an alternative source. When a model uses that timestamp to forecast course completion, the relationship is empirical and conditional on the sampled system. When an advisor uses the forecast to offer tutoring, the forecast becomes part of an intervention. When the institution uses the same prediction to restrict a student's access to opportunities, a different outcome, duty of justification, and ethical problem arise. Scientific evaluation must identify the specific decision and its consequences rather than treating all uses of an algorithm as one technology.
This monograph examines primary education, secondary schooling, university learning, and networked educational environments with explicit attention to the differences among their data-generating institutions. The center of the analysis is psychological: construct interpretation, motivation, self-regulated learning, feedback, trust, autonomy, social classification, and the effects of labels on expectations and conduct. Statistical learning, human–computer interaction, education policy, and data protection provide indispensable methodological and institutional context. The work is a critical narrative synthesis of identifiable studies, reviews, and public standards, with no new student data, original trial, or protocol-registered systematic-review claim.
Historical Formation of Educational Analytics
Educational decisions relied on records long before contemporary machine learning. Attendance logs, examinations, student files, and teacher judgments already created institutional data about learners. The psychometric tradition supplied methods for investigating whether such observations support particular interpretations. Cronbach & Meehl (1955) established construct validity as a theory-testing problem: an observed score becomes evidence about a construct through a network of justified relations and predictions. Messick (1989) connected validation to interpretation and consequences. The AERA, APA, & NCME (2014) further locate validity in evidence supporting proposed score interpretations and uses. These traditions remain fundamental when the observed variable is a log-derived feature or a model-produced probability instead of a conventional test score.
The growth of learning management systems introduced high-frequency traces into educational analysis: page views, assignment attempts, forum contributions, quiz responses, resource access, help requests, and temporal sequences. This expansion altered the granularity and apparent immediacy of monitoring. It also introduced dependencies on platform design and institutional requirements. A required button click and an optional forum contribution have different meanings. A course that uses short digital exercises cannot be compared directly with a studio course whose substantive work occurs offline. Dataset size increased faster than the availability of theoretical guarantees that every recorded event represented learning.
The first International Conference on Learning Analytics and Knowledge was held in 2011, marking an institutional milestone in the emergence of learning analytics as an interdisciplinary research field. Early scholarship explored large educational datasets, social networks, predictive models, and visual interfaces. Verbert et al. (2013) examined learning analytics dashboards as applications through which traces become legible to students and educators. Gašević et al. (2015) argued that analysis must remain accountable to learning rather than treating computational performance as its own educational endpoint. Wise & Shaffer (2015) likewise articulated the role of learning theory in the interpretation of large educational datasets.
The field subsequently expanded from descriptions of learning patterns to prediction, adaptive feedback, personalized support, and institutional decision systems. Governance became increasingly central as predictive models affected access to services and as student data moved among schools, platforms, and vendors. Contemporary debates integrate psychometrics, causal inference, fairness, information law, and participatory design. The historical trajectory explains why a current learning analytics system must be evaluated as a joint educational and institutional process. Its scientific quality depends on what is being measured, what a forecast represents, and what the institution does after a forecast is produced.
Learning Theory as the Basis of Interpretation
An educational trace acquires psychological meaning through a specified theory of learning. Behavioral accounts can interpret repeated practice and task completion in relation to reinforcement and performance. Cognitive accounts distinguish access to information, encoding, retrieval, mental effort, and transfer. Sociocultural approaches emphasize participation, social tools, norms, and the distributed organization of activity. Self-regulated learning theory emphasizes learners' goal formation, monitoring, strategy selection, and adjustment over time. The same observable behavior may occupy different positions in these explanatory systems. Model specification must state which relation is being tested.
Zimmerman (2002) described self-regulated learning as a dynamic process involving forethought, performance control, and self-reflection. Digital records can reveal some behavioral correlates of these phases, such as practice scheduling or repeated consultation of feedback. They cannot directly reconstruct a student's private intentions. Self-reports may reveal motives and interpretations absent from activity logs, but they have their own recall and response biases. A multimethod design can compare these sources and ask whether they converge on an interpretable construct.
Engagement is particularly vulnerable to overextension. Behavioral engagement, emotional engagement, cognitive engagement, and social participation are related but distinct. A frequently logged-in student may be struggling, complying with a course requirement, or performing repeated superficial actions. A student who rarely opens the platform may be learning through textbooks, group discussions, or accessibility tools outside the logging system. The validity of a log-derived engagement score depends on context-specific evidence linking the chosen indicators to learning activity and subsequent outcomes. A correlation with a final grade does not establish that every dimension of engagement is being measured.
Feedback also requires theory. Information helps when recipients understand it, regard it as actionable, possess sufficient opportunity to respond, and can incorporate it into their learning regulation. An analytics dashboard displaying percentile rank may redirect attention toward social comparison; a dashboard linking an error pattern to a specific practice strategy may support self-correction. The predicted psychological mechanisms differ. An intervention evaluated through Matcha et al. (2020) and related dashboard literature should therefore be classified by its learning-regulation functions rather than the mere presence of graphs.
From Observation to a Psychological Construct
Operationalization specifies which observations count as evidence for a psychological property. Student records often include administrative categories, prior grades, standardized scores, timestamps, submissions, online text, and interaction sequences. The measurement pathway requires a definition of the target construct, a justification for each proposed indicator, and evidence about missing or systematically excluded observations. A model predicting degree completion has an observable administrative target; a claim that it identifies low motivation is a stronger psychological interpretation that requires independent construct validation.
Construct contamination arises when a variable mixes the intended phenomenon with irrelevant conditions. Device availability, broadband access, work schedules, disability-related accommodations, platform navigation, and instructor reporting practices can affect counts of clicks and messages. Construct underrepresentation occurs when a target is only partially captured, such as measuring mathematical competence through timed online questions without examining reasoning or participation in collaborative work. Both problems can generate seemingly precise numbers whose interpretive scope is narrower than their label.
The timing of observation also matters. A trace recorded during the first two weeks of a course may be a predictor of later completion; an event recorded after a student has withdrawn cannot validly be treated as a prewithdrawal warning. A learning platform redesign can change what a click means from one semester to another. A teacher who migrates assignments from the LMS to a shared document changes the observation system while students' underlying learning might remain stable. The construction of valid analytic variables demands a documented relation among time, interface, population, and educational task.
Missing data are informative when their occurrence is related to opportunity or learning conditions. A student using an assistive device, studying offline, or declining an optional learning application may generate different data volumes. Complete-case analysis can disproportionately exclude such learners. Imputation procedures also encode assumptions about why observations are missing. Analysts should report completeness by relevant groups, sensitivity to assumptions, and the relation between missingness and both the educational outcome and the data-collection process. An apparently representative dataset must be tested against the actual population eligible for the resulting decisions.
Prediction, Explanation, and Educational Intervention
Prediction estimates an outcome under a defined data-generating regime. Explanation identifies relations that account for a phenomenon, while a causal claim describes how an outcome would change under an intervention. These are distinguishable research questions with different evidence requirements. A pattern associating irregular attendance with failure can be useful for forecasting; the same pattern does not establish that enforcing platform attendance will improve mastery. A predictive model may use proxies for unmeasured social conditions that are causes of both the trace and the educational outcome.
The transition from a risk estimate to an intervention introduces a decision policy. A university may offer mentoring to students with a forecasted low probability of completion. The value of that decision depends on prediction errors, intervention benefit, staff capacity, the costs imposed on recipients, and whether a student can decline or contest an intervention. The institution should therefore specify the eligible population, the threshold, the offered support, and the outcome before deploying the analytic system. AUC or classification accuracy alone cannot determine that policy's educational value.
Feedback changes future data. Once learners receive risk alerts, some change their behavior; once advisors respond to risk predictions, subsequent outcomes reflect the intervention as well as the student's prior circumstances. Models trained on historical outcomes may become miscalibrated when institutional support changes. Moreover, a prediction can affect teacher expectations or student identity, creating psychological effects that alter the outcome being predicted. Longitudinal evaluation must trace these feedback loops and measure whether favorable outcomes arise from useful support, changed recording practices, or unequal allocation of institutional attention.
The central scientific unit is therefore a decision pathway: specified population, justified construct, recorded data, predictive model, decision rule, communicating actor, intervention, student interpretation, and observed outcome. Each component needs its own measurement and audit criteria. This structure permits scientific comparison across institutions while preserving differences in curriculum, demographics, resources, and legal settings.
Data Architectures and the Educational Trace
An educational data architecture specifies events, identities, timeframes, relationships, and authorized purposes. Learning management systems usually generate event logs whose fields may include a learner identifier, event type, resource identifier, timestamp, course identifier, and contextual metadata. Assessment systems add responses, grading rubrics, feedback, item-level scores, and revision histories. Student information systems add enrollment, course history, program status, and institutional milestones. The scientific problem is the integration of records that were originally created for different organizational purposes.
Data integration is a substantive research decision. Course identifiers change, students transfer institutions, and instructional activities move between physical and digital environments. A learner may appear under several identifiers or participate through a shared device. Course activity may be stored at the class level while the intervention is assigned at the student level. Linkage errors can create spurious sequences, and missing links can split a continuous educational career into apparent discontinuities. Every proposed predictive study needs a provenance table describing record origin, collection mechanism, data permissions, temporal coverage, available identifiers, and exclusion rules.
An event trace and an educational event are different objects. A platform can record a file opening, yet the relevant educational process may involve several minutes of reading, a discussion, or subsequent application. Time-on-task estimates derived from page activity often assume that an open window represents attention. This assumption requires validation against observation, experience sampling, or independently measured learning outcomes. Automated video attention or emotion inference introduces further construct and privacy problems. The presence of a sensor or event field does not by itself confer validity on a psychological interpretation.
Gašević et al. (2015) identify the need to situate patterns in learning processes. Context-sensitive modeling requires representations of assignment difficulty, assessment timing, access conditions, teacher intervention, and curriculum. It also requires a distinction between institutional records that are essential to a support function and potentially intrusive records gathered merely because collection is technically possible. Data minimization is methodologically valuable: it forces investigators to explain which information meaningfully improves a specified decision and what additional uncertainty remains without it.
Feature Engineering, Sequencing, and Temporal Leakage
Feature engineering converts raw records into variables: number of attempts, fraction of completed tasks, delays between instruction and submission, regularity of study intervals, distribution of quiz errors, or network patterns of collaboration. These variables can represent educational mechanisms when the transformation is justified. The substantive meaning of an "attempt" depends on assessment design; a student allowed unlimited low-stakes retries will display different event counts from a student whose course permits a single attempt. Comparisons must incorporate these policy differences before attributing behavioral variation to the learner.
Sequence models attempt to preserve the order of actions. A learner's repeated revision after feedback may indicate productive strategy adjustment, while repeated submission without engagement with feedback may reflect an alternate process. The interpretation changes according to the content and timing of the feedback and the student's previous knowledge. State-transition models and process mining can generate hypotheses about learning pathways, but a sequence of interface events is a partial observation of a broader activity system. Researchers should avoid equating platform pathways with all meaningful routes to competence.
Temporal leakage occurs when information available only after a prediction point is included among features used to represent a forecast made earlier. Examples include a final grade, the eventual withdrawal date, later advisor notes, or an end-of-term activity count used to predict first-month risk. Leakage can also arise when data from the same student or cohort appear in both training and validation partitions, especially with repeated student records. Random record splitting may produce optimistic results if the intended application is prediction for a new cohort, teacher, program, or institution.
Validation should therefore replicate the moment at which a real decision would occur. A first-week alert needs features that existed by the end of the first week, and it must predict an outcome not yet observed. Temporal holdouts, institution-level validation, and nested model selection can test performance in settings closer to deployment. Sensitivity analyses can examine whether features remain predictive under course redesign, changing student composition, or the addition of support services. A claim of out-of-sample performance is meaningful only when the held-out observations match the anticipated generalization problem.
Predictive Models and Baselines
Educational prediction commonly uses logistic regression, regularized regression, decision trees, random forests, gradient-boosted models, and neural sequence architectures. A simple regression model can make relationships easier to inspect; a flexible ensemble can capture nonlinearity and interactions; a sequential architecture can exploit temporal patterns. Model choice depends on sample size, outcome definition, missingness, portability, interpretability, operational cost, and the consequences of error. Complexity is a hypothesis about useful structure, not a presumption of educational superiority.
A defensible study compares models against meaningful baselines. An institution might already know that prior course completion, early assignment performance, and current registration status predict subsequent outcomes. A new model should establish incremental value over those existing signals, including the cost of obtaining or linking additional data. A high headline accuracy in an imbalanced course-completion dataset can reflect the majority outcome. A model that predicts "completion" for almost everyone could look accurate while failing to identify students who would benefit from support.
The outcome should be articulated in operational terms: completion of a particular module by a particular date, persistence into the next term, passing a defined examination, or independently assessed mastery. Composite risk labels require transparent construction. Historical "success" often contains institutional policy, grading practices, financial constraints, and prior allocation decisions. These factors may be legitimate predictors of the administrative outcome while complicating the claim that the label measures academic capacity. Selection of the prediction target is an educational and ethical choice preceding model fitting.
Feature importance also requires careful interpretation. In correlated data, changing one feature can redistribute apparent importance to another. Permutation importance, partial dependence, and local explanation methods depend on the fitted model and the distribution of observations. They show how a statistical function uses information rather than establishing which educational intervention would change the outcome. For student support, causal studies of the available interventions are essential companions to predictive analysis.
Discrimination, Calibration, and Uncertainty
Discrimination evaluates how well a score separates cases with different observed outcomes. Receiver operating characteristic area under the curve and precision–recall analysis summarize different aspects of ranking ability, and their interpretation depends on the prevalence of the outcome and the action under consideration. Calibration examines whether predicted probabilities agree with observed frequencies. If a model assigns a probability of completion around 0.70 to a group of students, calibration asks whether completion occurs at approximately that frequency within the defined population. A well-ranked but poorly calibrated model can generate systematically misleading risk estimates.
An educational intervention policy needs probabilities and consequences, not only ranks. Suppose one cohort has a low failure rate and a second cohort has a higher rate because of institutional barriers. The same cutoff can produce different positive predictive values across cohorts. An apparently small false-positive rate may still create many unnecessary intrusive contacts when the event of interest is uncommon. Conversely, false negatives may leave students without timely support. Reporting should include sensitivity, specificity, precision, calibration curves, prevalence, uncertainty, and the decision threshold relevant to the actual institutional action.
Uncertainty arises from measurement, sample selection, model estimation, drift, and the irreducible variation of student circumstances. Confidence intervals around evaluation metrics can be wide for small subgroups; estimates in a new institution may differ substantially from the training institution. Analysts should quantify uncertainty under the conditions of deployment and explain the limits of individual forecasts. Presenting a 63% predicted probability as a fixed personal attribute converts a contingent population-based estimate into an unjustified judgment about a learner.
Transportability is a major research constraint. A predictor developed from one university's LMS, admission procedures, course structure, and support services may not generalize to a community college or a school district. Curriculum reform and platform changes can shift the predictor-outcome relation. Viberg et al. (2018) reviewed the higher education learning analytics landscape and highlighted the field's uneven development and limited evidence for learning improvement. Model updating, external validation, monitoring, and explicit local calibration are integral to scientific deployment rather than peripheral maintenance tasks.
Model Interpretability and Meaningful Human Oversight
Interpretability has several levels. A technical analyst may need access to model coefficients, validation reports, training provenance, subgroup metrics, and the code constructing features. An advisor needs an intelligible account of what a risk score means, which observations informed it, what decisions are permitted, and what responses are supported by evidence. A learner needs to understand how information about them is being used, what can be corrected, and how a significant decision may be questioned. One explanation interface cannot satisfy all these audiences automatically.
A human review step is meaningful when the reviewer has information, authority, time, and institutional permission to alter the outcome. Requiring a teacher to click "approve" on a model's recommendation creates an audit record but may add little independent judgment. Human reviewers can exhibit automation bias, anchoring, and selective attention to algorithmic labels. The scientific evaluation should compare decisions with and without the model, document when reviewers disagree, and examine whether overrides improve or worsen outcomes for different students.
Holstein et al. (2019) investigated teacher–AI complementarity in a real-time classroom orchestration tool, illustrating that effective support emerges from the relationship among analytics, instructional practice, and teacher knowledge. Their work supports an implementation research program attentive to workflow, actionable observations, and the teacher's informational needs. It does not establish that human presence alone makes an algorithmic decision just or accurate. The reviewer must have a genuine pathway for investigating local circumstances that the model cannot observe.
The strongest design begins with the educational action. Researchers identify a decision that qualified educators can legitimately make, determine which validated information would improve it, specify how uncertainty and disagreement are presented, and evaluate the resulting process. Explanations serve inquiry, accountability, and contestability. They become meaningful when students and staff can use them to correct data, modify decisions, and improve educational opportunities.
Data Quality, Documentation, and Reproducibility
A reproducible analytics study needs a data dictionary, data provenance statement, eligibility criteria, outcome construction, missing-data protocol, code or sufficient computational specification, validation design, and an account of changes made during model development. Researchers should distinguish exploratory analyses from preplanned evaluation, document candidate features tested, and report performance with uncertainty. Where student privacy prevents release of row-level records, synthetic examples, feature schemas, code, controlled research access, and audited aggregate outputs can support partial reproducibility.
Dataset versions matter. A vendor may alter an event taxonomy, change default interface behavior, or silently revise a predicted-risk field. An institution may update grades after appeals or correct registration statuses. Training and evaluation must use documented snapshots. Model cards and data documentation can identify intended use, excluded use, development population, validation samples, group performance, drift triggers, and named owners of corrective action. These descriptions remain useful only when institutions update them as systems change.
The data-governance implications follow from the measurement problem. A record retained indefinitely creates additional possibilities for secondary inference; a dataset assembled for academic support may become attractive for disciplinary screening, security monitoring, admissions targeting, or commercial profiling. Purpose limitation and access controls protect the interpretive validity of the original project as well as students' interests. Good governance preserves the conditions under which evidence is collected, interpreted, and legitimately used.
Empirical Research and Intervention Evidence
Educational analytics evidence includes observational analysis of activity traces, out-of-time prediction studies, dashboard evaluations, randomized feedback interventions, and evaluations of institutional implementation. Each method answers a different scientific question. A retrospective relation between digital activity and achievement tests association; a temporal holdout tests forecasting in a new period; random assignment of access to feedback tests the effect of that offer under particular conditions. A forecast can be accurate without a demonstrated benefit from acting on it, while an intervention can improve outcomes even when the supporting algorithm has modest discrimination.
Cerezo et al. (2016) examined interactions with a learning management system and their relations to achievement in higher education. Their work exemplifies the extraction of educationally interpretable behavior patterns from course-specific records. Its generalization requires attention to assessment design, interaction opportunities, and the difference between online behavior and unobserved studying. Matcha et al. (2020) subsequently organized empirical dashboard research from the perspective of self-regulated learning, highlighting a substantive question that aggregate login counts cannot resolve: whether the interface contributes to planning, monitoring, strategic adjustment, and reflection.
Dashboards have multiple mechanisms. Student-facing systems may help learners understand progress and revise their study behavior. Teacher-facing systems may direct attention to a student who would otherwise receive little feedback. Advisor-facing systems may create new outreach priorities. A dashboard often arrives together with tutoring, coaching, or automated messaging, making component attribution difficult. Sønderlund et al. (2019) reviewed learning analytics interventions in higher education, providing important context for the variety of implementations that are grouped under a single label.
The quantitative synthesis by Gestiardi & Yuniawatika (2026) specifically considered learning analytics dashboards separable from intensive co-interventions. It identified 22 studies comprising 33,409 participants, pooling 21. Student-facing dashboards showed Hedges' g = 0.287 across 16 studies; teacher-facing dashboards showed g = 0.260 across five, with the latter estimate judged fragile by the authors. The pooled descriptive baseline was g = 0.276, 95% confidence interval [0.162, 0.390]. Statistical heterogeneity was substantial, I-squared = 79.2%, and the 95% prediction interval extended from -0.161 to 0.712. This difference between uncertainty about the average and dispersion among implementations is decisive for institutional action: the review supports a small average improvement across published contrasts but allows that a new implementation may produce no improvement.
The same review reported that including two dashboard systems confounded with additional supports increased the descriptive baseline from g = 0.276 to 0.461. This sensitivity result shows the importance of accurately defining the intervention. It also reported more consistent effects for academic performance and completion than for self-reported self-regulation, log-based engagement, or motivation. The estimates do not justify a universal prescription to deploy dashboards. They justify evaluating particular dashboards, for a specified audience and outcome, in the context where they will be used.
Systematic Reviews and the Interpretation of Heterogeneity
Viberg et al. (2018) reviewed the landscape of learning analytics in higher education and emphasized the need for stronger evidence that analytics improves educational practice. A field can generate numerous technically sophisticated models while remaining less developed in experimental assessment of beneficial outcomes. Model accuracy, research publication count, and an institution's implementation maturity represent different dimensions of achievement.
García-García et al. (2026) conducted a meta-review of systematic reviews of learning analytics and academic performance, using a search extending to May 2025. The review-of-reviews design evaluates strengths and weaknesses in the accumulated synthesis literature. It must account for overlap between the primary studies summarized in several reviews. Multiple reviews of the same trials do not create independent experimental replications.
Differences among courses, educational levels, teacher practices, digital platforms, and the form of intervention can explain some variation in reported effects. Selection also matters. Students who consult a dashboard frequently may already be more motivated, better prepared, or more confident. Evaluation should therefore report both the effect of being assigned a dashboard and the association between actually using it and subsequent outcomes. Random assignment helps identify the former; analyses of the latter need explicit assumptions because usage is self-selected.
Primary outcomes should be specified in advance. Assignment completion, course grades, independent tests of mastery, self-regulation, psychological well-being, and subsequent persistence are not interchangeable endpoints. Researchers should include negative and null results and evaluate the durability of any observed gains. Learning analytics can reshape what is observed: encouraging recorded activity may increase platform metrics even when independent learning remains unchanged. Independent assessment and delayed measurement can help identify this distinction.
Subgroup Validity and Algorithmic Fairness in Educational Research
A model trained to predict past educational outcomes may absorb historical inequalities in resources, access, grading, and opportunities. Baker & Hawn (2022) review mechanisms of algorithmic bias in education. Fairness evaluation must articulate the educational outcome, the protected or substantively relevant groups, the forecast horizon, the threshold, and the consequences of errors. The assertion that a classifier has equal overall accuracy provides little information about whether its false negatives deprive some learners of support.
Bird et al. (2025) examined racial bias in two models of community-college success, predicting course completion and degree completion. Their study found that a policy targeting support using the evaluated risk predictions could leave some marginal Black students without needed resources. The size of bias depended on the target and the choice of risk cutoff. This result links model diagnostics directly to resource allocation. A technically accurate ranking can still distribute an institution's limited support inequitably when the underlying data are less informative about particular student groups.
Herrmann & Weigert (2024) compared prediction performance for traditional and nontraditional entrants, illustrating the significance of educational pathways underrepresented in standard training samples. Opoku et al. (2025) investigated fairness–accuracy trade-offs in machine learning using the Open University Learning Analytics Dataset, showing that bias mitigation can improve certain group metrics while preserving acceptable performance in the studied setting. Results in a public benchmark require local validation before they support operational decisions in another institution.
Fairness-aware model development includes attention to subgroup calibration, differences in false-positive and false-negative rates, decision thresholds, data coverage, and the opportunities available after a risk flag. A support offer and an exclusionary decision carry different consequences even if both use the same score. Institutions should evaluate decisions, including their actual human implementation, rather than attaching a fairness label to a predictive algorithm in isolation.
Student Data Governance and Educational Responsibility
Student data governance establishes authority over collection, linkage, interpretation, retention, disclosure, and use of information about learners. Its basic unit is a purpose-bound decision process. A system built for academic support specifies the learner population, data required, actors authorized to view results, permitted interventions, safeguards, correction procedures, and conditions for retiring the system. Governance extends to vendors and subcontractors handling institutional information, and it includes predictive scores and intervention histories alongside the original records.
Pardo & Siemens (2014) developed ethical and privacy principles for learning analytics. Drachsler & Greller (2016) proposed the DELICATE checklist for trusted learning analytics, while Ifenthaler & Schumacher (2016) studied student perceptions of privacy principles. These contributions establish that data protection, transparent purposes, institutional responsibilities, and student understanding are features of the educational process.
The Jisc (2015) treats student benefit as an organizing purpose, describes responsibilities for analytics and interventions, and emphasizes transparency, valid interpretation, appropriate access, and student agency. It recommends supportive rather than punitive uses of AI-enabled predictions. The code advises against using learning analytics for assessment, grading, or profiling within its particular practice framework. Other educational assessments require their own psychometric and legal validation.
Institutional responsibility must be assigned to identifiable offices. Educational staff determine whether an interpretation is instructionally meaningful, technical teams document and validate models, privacy and security functions examine lawful processing, and an authorized decision maker evaluates intervention. Procurement includes vendor documentation, access control, retention, and termination arrangements. Outsourcing computation does not remove the institution's duty to know how its predictions influence student opportunities.
Purpose Limitation and Data Minimization
Collection begins with the educational question. More recorded events do not automatically yield a better understanding of learning. A platform may record page visits, exact submission times, device metadata, and communication patterns; each variable still needs an account of its validity and expected value. Monitoring outside required learning activities may impose privacy costs without improving the decision. Researchers should compare predictive validity with and without optional intrusive features.
Records gathered to offer tutoring may later attract requests for discipline, admissions screening, employee surveillance, or marketing. Such purposes change the legitimacy and meaning of data. A support-oriented risk estimate cannot automatically justify restrictions. Reuse demands specific legal authority, a valid interpretation, transparent purposes, and an examination of consequences.
Retention procedures are scientific and administrative matters. Longitudinal studies may need bounded historical records, while indefinite storage creates additional exposure. Institutions should document retention periods for source records, joined datasets, training snapshots, scores, and backups. Consent also requires context. Students may depend on institutional platforms and have little practical ability to refuse routine processing. Participation in research, legal authority to process academic records, and voluntary acceptance of an intervention are different decisions.
Privacy, Security, and Trust
Removing student names does not guarantee anonymity. Course combinations, schedules, small cohorts, and external information can enable reidentification. Joining administrative and behavioral records increases this risk. Pseudonymization generally still requires access control. Security measures include restrictive permissions, authentication, encryption, logging, retention limits, and clear incident-response procedures.
Model-derived classifications and recommendations also require protection. Advisors may need an actionable indication about a support request without access to unrelated student history. The learner's experience of monitoring is itself a research question: students can change their participation when they expect invisible scoring. Transparent communication must explain which records are used, why, who sees them, how to obtain correction, and the limits of inference. Trust measures complement, and cannot replace, independent tests of validity, fairness, and legitimate institutional use.
Comparative Legal Frameworks and Educational Rights
Educational data governance operates under laws specific to the institution, learner population, purpose, and jurisdiction. An international scientific account examines those frameworks as distinct regulatory arrangements and identifies general requirements of validity, accountability, and contestability. Statutory compliance cannot be inferred from a model's accuracy or from the institution's benevolent intentions.
Within the European Union, the General Data Protection Regulation establishes principles concerning lawful, fair, transparent, purpose-limited and proportionate personal-data processing. The European Commission (n.d.) describes protections against certain solely automated decisions with legal or comparably significant effects, with exceptions and safeguards determined by law. Relevant procedures can include human intervention, opportunities to express a point of view, and the ability to contest a decision. The application of Article 22 requires examination of the actual degree of automation and significance of the result.
The European Parliament & Council (2024), places specified educational uses in high-risk categories, including certain systems concerned with access, educational level, or evaluation. Its provisions have staggered dates of application and require current operational verification. Risk classification depends on intended use. A course-level statistics dashboard, an adaptive practice tool, and an automated decision about admission have different functions and potentially different legal positions.
The European Commission (2026) present a practical framework for ethical AI and data use in teaching. The guidance focuses substantially on primary and secondary settings and on educator competence. It complements research on the psychological effects of monitoring, explanations, student agency, and teacher judgment. International comparison should distinguish guidance documents, binding regulations, research standards, and institutional policies.
In the United States, the Family Educational Rights and Privacy Act establishes rules concerning access to and disclosure of education records by covered institutions. The U.S. Department of Education (n.d.) discusses disclosure recording, qualifying exceptions, and institutional responsibilities. Certain service-provider arrangements may be permitted under the school-official exception when its conditions are satisfied. This is a conditional legal arrangement, not a general authorization for unlimited secondary data use. Other national and regional laws address children's information, disability access, and institutional records under their respective scopes.
Contestability and Meaningful Review
A contestable educational decision provides a route to understand that analytics influenced the result, inspect relevant data, correct errors, contribute contextual information, and seek review by someone with authority to change the outcome. A merely nominal human reviewer may reproduce the algorithm's decision through automation bias. Meaningful oversight requires sufficient time, access to relevant evidence, and procedures allowing an actual revision of the result.
Students need accessible explanations of how educational evidence is interpreted. The explanation should describe the specific outcome forecast, the uncertainty involved, the legitimate use of the prediction, and available actions. A list of influential features offers only a partial model explanation. It does not necessarily show the causes of a student's difficulties or identify which intervention would help. A person with authority to decide must investigate student circumstances beyond the dataset when they materially affect a consequential decision.
Appeals provide additional scientific evidence. Correction requests reveal errors in identity linkage, timestamp interpretation, missing accommodations, or historical records. Institutional audits should aggregate error categories, rates of successful appeal, differential access to remedies, and subsequent educational consequences. Complaint data alone may be misleading because students differ in knowledge of procedures, confidence in institutional authority, and resources needed to contest a classification.
Procurement, Vendors, and Model Accountability
A school adopting a proprietary analytics system should define its decision purpose before procurement. Vendors should document data origin, intended application, model training populations, performance, uncertainty, known limits, accessibility, and change-management procedures. Institutions need to know whether student data are used to train models serving unrelated customers and whether the provider can delete the data under an enforceable agreement.
Opacity in proprietary systems creates a verification challenge. An institution may be unable to inspect all implementation details while still being responsible for decisions made using the product. Independent performance tests, controlled access to documentation, contractual audit rights, and contestable procedures can reduce some of that risk. An institution should retain an option to suspend use when validation fails or the vendor materially changes the model without appropriate review.
The scientific standard of governance is sustained evidence that a particular decision process is justified, effective, and revisable. Regulation can establish enforceable minimum duties, while research investigates construct validity, benefits and harms, mechanisms of behavioral response, and variation among student groups. These complementary activities support education institutions acting responsibly with information about learners.
Causal Evaluation of Algorithmic Educational Decisions
Predictive validity establishes how an estimate relates to an outcome under specified conditions. Causal evaluation asks whether an analytics-informed action changes that outcome compared with a credible alternative. The distinction is central to educational governance. Students identified as at risk may improve because they receive additional mentoring, because their instructors modify teaching, or because they obtain financial assistance. A risk score cannot establish which of these interventions causes improvement.
The evaluation unit should match the action. A student-level notification may be randomized among eligible learners, while an advisor-workflow change may require randomization by advisor, course, or institution to avoid contamination. Cluster-randomized designs must account for intraclass correlation and differences in baseline conditions. Stepped-wedge designs may be useful when an institution plans sequential rollout, but secular trends and allocation timing still require scrutiny. For interventions that cannot be randomized, credible quasi-experimental approaches may exploit known policy changes, eligibility thresholds, or longitudinal comparison structures, with clearly articulated identification assumptions.
A well-designed evaluation distinguishes assignment to the intervention from actual engagement with it. The intention-to-treat effect estimates the consequences of offering the analytics-informed intervention, including variable uptake. A per-protocol or as-treated comparison can be confounded because motivated students and staff are more likely to engage. Intermediate outcomes, such as viewing a dashboard or scheduling an advising appointment, inform mechanisms but should not replace independent measures of learning and persistence.
Equity effects also require causal evaluation. A model may provide reasonable calibration across groups while allocating an intervention that helps some students more than others. Conversely, an intervention may improve overall persistence but widen inequality if access is uneven. Studies should estimate heterogeneous treatment effects where sample size and study design permit, report uncertainty, and avoid treating exploratory subgroup differences as established effects. The intervention's actual distribution and student uptake are part of the outcome.
Decision Thresholds and Allocation of Educational Support
A score is often transformed into an action by a threshold. The threshold determines who is contacted, invited to support, flagged for review, or subjected to an institutional consequence. In settings with limited advisor capacity, threshold choice becomes a resource-allocation policy. A mathematically optimized cutoff depends on the cost of erroneous referrals, the value of effective support, and the harms arising when students who need assistance are overlooked. These values require public institutional justification.
Different thresholds can generate different group disparities, as the community-college analysis by Bird et al. (2025) illustrates. The decision maker should inspect not only accuracy metrics but also expected support coverage, subgroup access, the costs of outreach, and the expected benefit from the offered service. If support is inexpensive, nonstigmatizing, and accessible, broader eligibility may be appropriate. If the intervention is intrusive or imposes restrictions, the evidentiary and procedural requirements are correspondingly more demanding.
Decision-curve or utility analyses can clarify trade-offs if the assumed costs and benefits are disclosed. However, the institution's goals may include learner autonomy, equity, protection from stigma, and trust, which cannot always be collapsed into a single numerical utility. A transparent policy should identify its nonnegotiable safeguards and explain how resource constraints are managed. Monitoring should assess the observed decisions after rollout, not solely the performance of a model before deployment.
Independent Verification and Replication
A robust scientific program requires multiple forms of replication. Computational reproduction tests whether the reported analysis can be rerun from the same specification and data. Internal validation examines stability within the development environment. Temporal validation tests later cohorts. External validation tests new institutions or populations. Conceptual replication investigates the same proposed learning mechanism through alternative measures or designs. Each addresses a different source of uncertainty.
A dataset repeatedly used in the research literature can make comparative modeling convenient while restricting the diversity of situations tested. Authors should report the overlap between training and test data, the method of hyperparameter selection, and any model tuning performed after inspecting benchmark results. Independent external validation requires a genuinely new setting and a predefined outcome. Prospective impact evaluation provides stronger evidence that decisions informed by the system are beneficial in use.
The reference standard for outcomes needs auditing. A student who is counted as having failed to persist may have transferred, taken a deliberate break, or moved into an alternative credential program. A student's low score may reflect inequitable assessment rather than weak mastery. Outcome misclassification affects model comparison, subgroup calibration, and the interpretation of an intervention. Reviewers should inspect the institution's administrative definitions rather than assuming that database labels are natural psychological categories.
An institution can adopt monitoring criteria for data drift, prediction calibration, support capacity, contested decisions, and student-reported harm. Changes in course delivery, regulations, student composition, or access to learning technology may invalidate assumptions used in training. A prospective audit schedule should specify who investigates these shifts, which failures require suspension, and what evidence permits revised deployment.
Evidence Quality and a Practical Assessment Matrix
The quality of evidence depends on the claim being made. A theory paper can clarify the construct and its plausible mechanisms. A cross-sectional study can show association under the measured conditions. A longitudinal study can establish temporal order but may still contain unmeasured confounding. A randomized intervention study can estimate the effect of a specified treatment, subject to its fidelity and generalization limits. A meta-analysis can summarize comparable effects but cannot repair fundamental bias in the underlying studies.
Each proposed educational deployment can be assessed across six questions: Is the measured construct valid? Does the model predict with acceptable calibration and transportability? Are subgroup errors and distributions of support acceptable? Is there evidence that a specific intervention benefits learners? Is the data practice lawful, proportionate, secure, and contestable? Can the institution maintain monitoring and remedy over time? A favorable answer in one domain cannot substitute for evidence in the others.
The most common scientific error is moving directly from retrospective model performance to a claim of improved education. The most common governance error is describing an algorithm as merely advisory while its output determines access to scarce resources. A complete assessment identifies who actually acts and what happens to students as a result. That makes the relationship between psychological science and institutional accountability measurable.
Developmental Differences and Educational Contexts
A predictive learning system deployed with young children has different psychological and institutional implications from one used with adult university students. Development affects reading comprehension, self-regulation, the interpretation of feedback, and the capacity to understand or contest a recommendation. The meaning of engagement indicators also varies with curriculum and age. Early-primary education often involves substantial face-to-face learning and teacher mediation outside a digital platform, while some university courses rely heavily on recorded asynchronous activity.
Children's dependence on schools and caregivers intensifies the importance of data minimization, accessible explanations, and proportionate institutional responses. A child who has little opportunity to understand an automated classification may experience it through altered teacher expectations or restricted access to enrichment. Researchers should assess whether educators and families receive a valid account of uncertainty and whether the student can present relevant circumstances. Safeguards must be sensitive to age, disability, cultural context, and the seriousness of any resulting decision.
School attendance and student participation are shaped by transportation, family responsibilities, illness, displacement, and access to technology. Predictors trained on online platform usage can consequently mix learning with environmental constraints. Where students use accessible formats or receive accommodations, recording practices can differ even when instructional goals are equivalent. Inferences about motivation, ability, or responsibility from these records need explicit evidence of measurement comparability.
A cross-institutional model should document which settings contribute to training and independent validation. Evidence collected from a single national education system or an online degree program has limited direct reach into primary classrooms or vocational training. When studies report a general effect of learning analytics, readers need population, period, instructional setting, outcome measure, and intervention detail before applying the result to their own institutions.
Disability, Accessibility, and Inclusive Interpretation
Digital education presents both opportunities for accessible feedback and risks from interfaces that register disability-related behavior as anomalous. Screen readers, alternate input devices, assistive communication systems, extended time, and adjusted assignment schedules can change trace patterns. A model should not assume that the average platform event sequence represents a universal learning pathway. Accessibility testing is therefore part of measurement validation and not only a final user-interface check.
A model trained on conventional testing conditions may classify an accommodation-related pattern as poor performance. Fair evaluation must determine whether the intended target is independent mastery, demonstrated performance under a legitimate accommodation, or a different administrative outcome. A supported student's outcome is partly shaped by the quality of institutional accommodations. Depriving the student of additional opportunities because a model treats the accommodation as risk can create a recursive disadvantage.
Dashboard interfaces should permit keyboard navigation, appropriate contrast, clear labels, screen-reader interpretation, and alternatives to color-only signals. They should communicate uncertainty in language appropriate to the intended audience. In a student-facing interface, a risk estimate without an actionable explanation can be particularly difficult to interpret. Inclusive design can be evaluated through usability studies involving learners with differing needs, followed by independent assessment of whether the interface actually changes learning strategies or outcomes.
A fairness audit should report how disability or accommodation information was collected and protected, whether group-level comparisons are statistically reliable, and how decisions handle students whose experiences are absent from the training set. Small subgroups can make estimates noisy and increase reidentification risk. Sometimes a qualitative investigation of errors and a carefully designed case review will be more informative than a falsely precise numerical disparity estimate.
Cultural, Linguistic, and Institutional Transportability
Educational institutions differ in cultural expectations about participation, help seeking, interaction with teachers, and visible public discussion. A model treating silence in a forum as disengagement may fail in settings where learners participate through other channels or where public posting is discouraged. Language proficiency can affect reading time, response patterns, and the precision of automated analysis. These differences require contextual learning theory and empirical measurement invariance tests.
Internationally deployed systems must address different calendars, grading scales, definitions of completion, modes of enrollment, and legal duties. A variable labeled "withdrawal" can refer to different institutional events. Even where the same target label exists, the opportunities and costs associated with it may differ. The scientific response is to specify the target process, not merely to translate field names and reuse a trained model.
Educational inequality also shapes data availability. A learner with limited connectivity may study extensively through downloaded materials and generate fewer interaction records. Students with employment and caregiving duties may use irregular study schedules that do not align with the behavioral norms of full-time campus learners. Models trained to interpret regular daytime platform activity as desirable can transform differences in opportunity into apparent differences in ability. Performance metrics need to be read in the context of the data-generating environment.
Cross-cultural comparison can strengthen research when investigators use common construct definitions, carefully aligned outcomes, and independent samples. It can also uncover mechanisms that appear stable despite institutional variation, such as the benefits of timely interpretable feedback under particular conditions. The relevant ambition is transportable causal knowledge with documented conditions, not the assumption that a single risk score means the same thing everywhere.
Faculty, Advisors, and Institutional Workload
Educational analytics affects teachers and advisors as well as students. A prediction system may create more requests for outreach than personnel can address, fragment attention, or shift professional judgment toward recorded indicators. It can also help educators identify misconceptions and prioritize useful conversations. The organizational effect depends on staffing, authority, training, system reliability, and how feedback is integrated into real work.
Research should examine the full workflow: alert generation, prioritization, advisor contact, student response, follow-up support, and documentation. A notification that no one reads has little educational effect. A notification that overburdens advisors can displace other valuable activity. Studies need data about staff response times, intervention fidelity, opportunity costs, and the experiences of learners who receive no contact. The fact that a dashboard was technically available is insufficient evidence that it changed practice.
Teacher professional knowledge can correct an algorithm's incomplete representation of a learner, while algorithms can reveal patterns teachers have difficulty observing at scale. The teacher–AI complementarity investigated by Holstein et al. (2019) gives this relationship a concrete design research agenda. It also reinforces the need to test whether decision makers can understand, appropriately reject, and revise algorithmic recommendations. Staff capacity and accountability are part of the causal mechanism of an analytics-informed intervention.
Learning Analytics in the Artificial Era
The Artificial Era introduces educational environments in which artificial systems generate explanations, adapt exercises, assess drafts, recommend learning pathways, and participate in the production of intellectual work. Learning analytics increasingly draws on interaction traces from these systems. The scientific object remains the organization of learning and decisions, while the observable environment now includes contributions from learners, teachers, automated tutors, and collaborative artificial agents. This development raises questions about the provenance of work, independent competence, reliance on external cognitive resources, and the accountability of institutions that delegate instructional functions.
A text generated during a tutorial session may include student reasoning, model-produced suggestions, retrieved documents, and subsequent revision. A dataset that treats the entire conversation as the student's own cognitive output risks construct contamination. The analytical record needs to distinguish the learner's initial performance, assistance received, actions taken after assistance, and independent performance when assistance is withdrawn. An apparently successful assisted answer can coexist with limited durable learning. Conversely, thoughtful use of external support can facilitate understanding when the learner can later transfer relevant knowledge and skills.
The relevant psychological constructs include metacognition, calibration of confidence, strategy choice, help seeking, attention allocation, and transfer. Learning analytics can estimate how learners interact with artificial feedback, but the validity of such estimates depends on independent assessments. Evaluating only the fluency or correctness of assisted outputs measures a joint human–artificial process. Research on independent learning requires separate outcomes under defined assistance conditions. Studies can compare immediate assisted performance, delayed unassisted retention, near transfer, and far transfer in new tasks.
The UNESCO (2023) addresses the educational and governance implications of generative systems. The European Commission (2026) provide an additional institutional frame. These documents do not establish any single model of psychological development under AI assistance. They clarify why pedagogical design, data processing, appropriate oversight, and student rights must remain linked as artificial systems become integrated into education.
Human–Artificial Decision Partnership
A human–artificial educational partnership is an organized relationship among machine-generated analysis, qualified educational judgment, and the learner's agency. Its quality depends on how information is distributed and whether each participant has an appropriate role in the decision. Artificial prediction can help reveal patterns; teachers and students contribute knowledge of goals, circumstances, and the meaning of activity. The partnership becomes scientifically evaluable when the contributions and final outcomes can be separated and measured.
The presence of a human in the workflow does not guarantee independence of judgment. Reviewers may defer to apparently objective numerical classifications or resist accurate recommendations for reasons unrelated to evidence. Institutions should investigate both automation bias and inappropriate rejection. Experimental comparisons can examine whether decision makers receive a score, an explanation, uncertainty information, or alternative data, and whether those conditions change accuracy, resource allocation, and subsequent student outcomes. Effective oversight requires genuine authority to override and a procedure for learning from disagreements.
A useful analytical distinction is between a recommendation and an institutional decision. A system may recommend an additional practice problem based on a student's recent errors. It may also recommend moving a student into a restricted academic track. The evidentiary requirements and potential harms differ markedly. The first action can be low stakes, reversible, and immediately informative; the second may create durable effects on opportunity and identity. Institutions should scale oversight, validation, and rights protection to the actual consequences of the action.
Artificial systems also change the boundary between instructional data and private expression. A learner might disclose distress, family circumstances, or personal beliefs while conversing with an educational assistant. Such information may be irrelevant to assessing mastery and inappropriate for broad institutional analysis. Governance requires explicit policies for collection, retention, and access to sensitive conversational material. Scientific ambition does not justify converting every interaction into a permanent psychological profile.
Unresolved Problems of Measurement and Agency
A major open question is the reliability of inferring self-regulation from increasingly complex human–artificial interactions. A learner who repeatedly requests explanations may be actively reflecting or may be substituting machine responses for independent effort. A model trained on conversation length or help-request frequency may misclassify both. Researchers need longitudinal multimethod designs incorporating performance, interviews, process tracing, and tasks that distinguish assistance from transferable competence.
Another problem is the effect of personalized risk feedback on learner identity. If students encounter persistent labels such as "likely to fail," their interpretation can affect perceived self-efficacy and institutional belonging. Conversely, an explanation that preserves uncertainty and offers actionable support may encourage adaptive response. These hypotheses need experiments that compare communication designs and measure psychological outcomes alongside academic performance. The mere availability of a probability estimate does not determine the effect of its communication.
The growth of generative systems also creates new sources of drift. Model updates can change the difficulty, style, and accuracy of feedback and thus alter the traces used by predictive learning analytics. When two cohorts use different versions of an assistant, their interaction records may not be comparable. Research should log relevant model versions and assistance policies, evaluate changes in independent outcomes, and revise measurement models when the digital environment changes.
Artificial cognition and artificial agency are emerging objects of interdisciplinary inquiry. Their psychological relevance to education should be expressed through observable behaviors and clearly defined hypotheses. A system's ability to recommend an instructional action does not establish that it possesses human-like understanding or motivation. Scientific evaluation focuses on its operational performance, the human responses it elicits, and the learning consequences of the joint system. The conceptual horizon of Artificial Era provides a broader frame for inquiry without replacing the evidentiary requirements of educational psychology.
Research Priorities for Human–Artificial Educational Environments
Future research should develop transparent datasets that distinguish independent student work from assistance; conduct randomized tests of feedback design; evaluate long-term retention and transfer; validate predictions across educational cultures; and examine agency, trust, and equity as primary outcomes. Data governance research should compare student-controlled data access, institutional oversight structures, vendor contracts, and contestability procedures. Causal evaluations should test the educational benefit of decisions rather than only the quality of automated forecasts.
The most promising unit of analysis is a complete human–artificial educational decision pathway. It includes the original task, the human and artificial contributions, the measured construct, the model's forecast and uncertainty, the educator's decision, the learner's response, and independent outcomes. Each component can be documented, challenged, and improved. Such research supports a cumulative science of effective educational partnership while maintaining institutional accountability for decisions affecting learners.
Critical Scientific Discussion
Learning analytics research has developed strong methods for describing educational behavior and predicting institutional outcomes. Its most consequential questions concern how forecasts are interpreted and used. Retrospective studies can establish that particular traces contain information about future outcomes. Their generalization depends on population, curriculum, interface, and decision environment. Construct validation is necessary before a trace becomes a claim about motivation or learning; a credible intervention evaluation is necessary before an analytics-informed practice is credited with improving education.
The research literature provides evidence of heterogeneous and often modest benefits from dashboards. The Gestiardi & Yuniawatika (2026) meta-analysis separates more isolated dashboard effects from bundled interventions and presents a prediction interval that crosses zero. A positive pooled mean does not establish a positive result in every course. Institutions need local pilots, specified educational outcomes, suitable comparisons, and continuing evaluation after deployment.
Evidence on algorithmic bias identifies a further limitation. Educational outcomes result from instruction, resources, policies, and social conditions as well as individual action. Models trained to reproduce historical outcomes can preserve patterns of unequal opportunity. Bird et al. (2025) demonstrate that risk thresholds can distribute access to support unequally. Baker & Hawn (2022) describe multiple origins of algorithmic bias in education, and Opoku et al. (2025) investigate fairness–accuracy relationships in student prediction models.
The institutional question is what happens after prediction. A model influences education through a decision, communication, or intervention. A risk label may prompt supportive inquiry, substitute for individualized attention, or determine access to an opportunity. Its consequences depend on the action, the student's response, and the alternatives available. Technical evaluation must therefore be joined to causal and institutional assessment.
Data governance supports scientific reproducibility. Documented provenance, legitimate access, controlled retention, correction processes, and records of decisions make it possible to reconstruct how forecasts were produced and used. Uncontrolled secondary processing, undocumented model changes, and missing audit records compromise both student rights and the integrity of claims. The scientific unit of accountability is the entire decision pathway from observed learning activity to consequential institutional action.
Open Research Questions
The first open question concerns the stability of psychological constructs inferred from digital behavior. Researchers must test whether indicators of engagement, help seeking, self-regulation, and mastery retain meaning across different ages, curricula, accessibility arrangements, and platforms. Independent assessment and student accounts are valuable because interface traces capture only part of educational activity.
The second concerns causal effects. Studies should compare analytics-informed decisions against credible alternatives, differentiate dashboards from co-interventions, and measure durable knowledge, student agency, equity, and staff workload. The third concerns transportability: models need independent temporal and institutional validation, monitoring for calibration drift, and explicit suspension criteria when underlying conditions change.
The fourth concerns governance effectiveness. Adoption of a written code does not by itself demonstrate that students can correct records or contest adverse outcomes. Comparative institutional research should study actual rights awareness, access to remedy, fairness of resource allocation, and vendor accountability. The fifth concerns human–artificial educational environments, where analysts must distinguish independently demonstrated understanding from machine-assisted output and assess the psychological consequences of long-term cognitive partnership.
Conclusion
Learning Analytics, Algorithmic Educational Decisions, and Student Data Governance form a coherent scientific object because educational data acquire practical consequences through interpretation and institutional action. Learning analytics converts observations into claims about learners, learning processes, and educational environments. Statistical models estimate conditional outcomes or classify observed patterns. Educational decision systems allocate attention, feedback, support, and opportunities. Governance defines the authorized purposes of these activities, the protection and interpretation of student records, and the processes through which consequential decisions remain accountable and correctable.
The psychological foundations of this field concern construct validity, self-regulation, motivation, feedback, autonomy, and the effects of institutional labels. Digital records offer an extensive view of interactions with educational systems, while leaving important aspects of understanding, circumstances, and independent performance unobserved. Sound inference requires explicit measurement theory. A frequency count of platform activity is an observation. Its interpretation as engagement, effort, or competence is a scientific claim requiring independent justification.
The empirical evidence supports conditional educational benefits from certain dashboards and analytics-informed interventions. The contemporary synthesis of controlled dashboard studies identifies a modest positive average alongside substantial variation across implementations. It establishes the value of testing specific systems under specific conditions rather than presuming that digital monitoring improves instruction. Original research on fairness shows that predictive error can alter which students receive support and that this distribution may vary sharply with decision thresholds, target definitions, and institutional context.
A complete evaluation connects measurement, predictive calibration, the decision policy, human review, and subsequent educational outcomes. Its scientific purpose is to determine whether the system improves relevant learning and opportunity compared with appropriate alternatives. Its governance purpose is to establish that the records and conclusions used for those decisions are legitimate, proportionate, secure, explainable, and open to correction. Both purposes require continued monitoring after deployment because curricula, student populations, platforms, institutional policies, and artificial systems evolve.
The Artificial Era extends this research into educational systems where human and artificial agents jointly produce explanations, assessments, and decisions. The central methodological requirement is to distinguish assisted output from independently demonstrated understanding and to examine what students learn, retain, and transfer when assistance changes. Artificially mediated educational decisions introduce new sources of uncertainty and data provenance concerns while expanding opportunities for timely feedback and collaborative inquiry.
The enduring scientific task is to establish the conditions under which learning analytics provides valid knowledge and supports beneficial, equitable, and revisable educational action. The unit of explanation is the complete pathway from observed activity to meaningful educational consequence. Its quality is determined by verified constructs, credible research designs, accountable institutional practice, and the capacity of students and educators to understand and revise decisions affecting learning.
References
- American Educational Research Association, American Psychological Association, & National Council on Measurement in Education. (2014). Standards for educational and psychological testing. American Educational Research Association. https://ncme.org/resources/books/testing-standards/
- Baker, R. S., & Hawn, A. (2022). Algorithmic bias in education. International Journal of Artificial Intelligence in Education, 32(4), 1052–1092. https://doi.org/10.1007/s40593-021-00285-9
- Bird, K. A., Castleman, B. L., & Song, Y. (2025). Are algorithms biased in education? Exploring racial bias in predicting community college student success. Journal of Policy Analysis and Management, 44(2), 379–402. https://doi.org/10.1002/pam.22569
- Cerezo, R., Sánchez-Santillán, M., Paule-Ruiz, M. P., & Núñez, J. C. (2016). Students' LMS interaction patterns and their relationship with achievement: A case study in higher education. Computers & Education, 96, 42–54. https://doi.org/10.1016/j.compedu.2016.02.006
- Cronbach, L. J., & Meehl, P. E. (1955). Construct validity in psychological tests. Psychological Bulletin, 52(4), 281–302. https://doi.org/10.1037/h0040957
- Drachsler, H., & Greller, W. (2016). Privacy and analytics: It's a DELICATE issue: A checklist for trusted learning analytics. In Proceedings of the Sixth International Conference on Learning Analytics & Knowledge (pp. 89–98). Association for Computing Machinery. https://doi.org/10.1145/2883851.2883893
- European Commission, Directorate-General for Education, Youth, Sport and Culture. (2026). Guidelines on the ethical use of artificial intelligence and data in teaching and learning for educators. Publications Office of the European Union. https://doi.org/10.2766/9548952
- European Commission. (n.d.). Information for individuals: Rights in relation to automated decision-making and profiling. https://commission.europa.eu/law/law-topic/data-protection/information-individuals_en
- European Parliament & Council of the European Union. (2024). Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act). Official Journal of the European Union. https://eur-lex.europa.eu/eli/reg/2024/1689
- García-García, F. J., Gómez-Núñez, M. I., & Molla-Esparza, C. (2026). Applications of learning analytics in the study of academic performance in higher education: A meta-review with an educational perspective. Higher Education. https://doi.org/10.1007/s10734-026-01709-y
- Gašević, D., Dawson, S., & Siemens, G. (2015). Let's not forget: Learning analytics are about learning. TechTrends, 59(1), 64–71. https://doi.org/10.1007/s11528-014-0822-x
- Gestiardi, R., & Yuniawatika. (2026). Learning analytics dashboards in Internet-enabled higher education: A meta-analysis of effects on student outcomes, self-regulated learning, and engagement. The Internet and Higher Education, Article 101119. https://doi.org/10.1016/j.iheduc.2026.101119
- Herrmann, L., & Weigert, J. (2024). AI-based prediction of academic success: Support for many, disadvantage for some? Computers and Education: Artificial Intelligence, Article 100303. https://doi.org/10.1016/j.caeai.2024.100303
- Holstein, K., McLaren, B. M., & Aleven, V. (2019). Co-designing a real-time classroom orchestration tool to support teacher–AI complementarity. Journal of Learning Analytics, 6(2), 27–52. https://doi.org/10.18608/jla.2019.62.3
- Ifenthaler, D., & Schumacher, C. (2016). Student perceptions of privacy principles for learning analytics. Educational Technology Research and Development, 64(5), 923–938. https://doi.org/10.1007/s11423-016-9477-y
- Jisc. (2015; updated 2026). Code of practice for learning analytics. https://www.jisc.ac.uk/guides/code-of-practice-for-learning-analytics
- Matcha, W., Uzir, N. A., Gašević, D., & Pardo, A. (2020). A systematic review of empirical studies on learning analytics dashboards: A self-regulated learning perspective. IEEE Transactions on Learning Technologies, 13(2), 226–245. https://doi.org/10.1109/TLT.2019.2916802
- Messick, S. (1989). Meaning and values in test validation: The science and ethics of assessment. Educational Researcher, 18(2), 5–11. https://doi.org/10.3102/0013189X018002005
- Opoku, R. A., Pei, B., & Xing, W. (2025). Unveiling accuracy-fairness trade-offs: Investigating machine learning models in student performance prediction. Journal of Learning Analytics, 12(2), 125–139. https://eric.ed.gov/?id=EJ1483325
- Pardo, A., & Siemens, G. (2014). Ethical and privacy principles for learning analytics. British Journal of Educational Technology, 45(3), 438–450. https://doi.org/10.1111/bjet.12152
- Society for Learning Analytics Research. (2026). What is learning analytics? https://www.solaresearch.org/about/what-is-learning-analytics/
- Sønderlund, A. L., Hughes, E., & Smith, J. (2019). The efficacy of learning analytics interventions in higher education: A systematic review. British Journal of Educational Technology, 50(5), 2594–2618. https://doi.org/10.1111/bjet.12720
- U.S. Department of Education, Student Privacy Policy Office. (n.d.). Frequently asked questions: Protecting student privacy. https://studentprivacy.ed.gov/frequently-asked-questions
- UNESCO. (2023; updated 2026). Guidance for generative AI in education and research. https://www.unesco.org/en/articles/guidance-generative-ai-education-and-research?hub=66973
- Verbert, K., Duval, E., Klerkx, J., Govaerts, S., & Santos, J. L. (2013). Learning analytics dashboard applications. American Behavioral Scientist, 57(10), 1500–1509. https://doi.org/10.1177/0002764213479363
- Viberg, O., Hatakka, M., Bälter, O., & Mavroudi, A. (2018). The current landscape of learning analytics in higher education. Computers in Human Behavior, 89, 98–110. https://doi.org/10.1016/j.chb.2018.07.027
- Wise, A. F., & Shaffer, D. W. (2015). Why theory matters more than ever in the age of big data. Journal of Learning Analytics, 2(2), 5–13. https://doi.org/10.18608/jla.2015.22.2
- Zimmerman, B. J. (2002). Becoming a self-regulated learner: An overview. Theory Into Practice, 41(2), 64–70. https://doi.org/10.1207/s15430421tip4102_2