Search

Background: Computer-based assessment allows for the monitoring of reader behaviour. The identification of patterns in this behaviour can provide insights that may be useful in informing educational interventions. Objectives: Our study aims to explore what different patterns of reading activity exist, and investigates their interpretation and consistency across different task sets (units), countries, and languages. Three patterns were expected: on-task, exploring and disengaged. Methods: Using log data from the PISA 2012 digital reading assessment (9226 students from seven countries), we conducted hierarchical cluster analyses with typical process indicators of digital reading assessments. We identified different patterns and explored whether they remained consistent across different units. To validate the interpretation of the identified patterns, we examined their relationship to performance and student characteristics (gender, socio-economic status, print reading skills). Results and Conclusions: The results indicate a small number of transnational clusters, with unit-specific differences. Cluster interpretation is supported by associations with student characteristics--for example, students with low print reading skills were more likely to show a disengaged pattern than proficient readers. Exploring behaviour tended to be exhibited only once across the three units: It occurred in the first unit for proficient readers and in later units for less skilled readers. Major Takeaways: Behavioural patterns can be identified in digital reading tasks that may prove useful for educational monitoring and intervention. Although task situations are designed to evoke certain behaviours, the interpretation of observed behavioural patterns requires validation based on task requirements, assessment context and relationships to other available information.

The NAEP EDM Competition required participants to predict efficient test-taking behavior based on log data. This paper describes our top-down approach for engineering features by means of psychometric modeling, aiming at machine learning for the predictive classification task. For feature engineering, we employed, among others, the Log-Normal Response Time Model for estimating latent person speed, and the Generalized Partial Credit Model for estimating latent person ability. Additionally, we adopted an "n"-gram feature approach for event sequences. Furthermore, instead of using the provided binary target label, we distinguished inefficient test takers who were going too fast and those who were going too slow for training a multi-label classifier. Our best-performing ensemble classifier comprised three sets of low-dimensional classifiers, dominated by test-taker speed. While our classifier reached moderate performance, relative to the competition leaderboard, our approach makes two important contributions. First, we show how classifiers that contain features engineered through literature-derived domain knowledge can provide meaningful predictions if results can be contextualized to test administrators who wish to intervene or take action. Second, our re-engineering of test scores enabled us to incorporate person ability into the models. However, ability was hardly predictive of efficient behavior, leading to the conclusion that the target label's validity needs to be questioned. Beyond competition-related findings, we furthermore report a state sequence analysis for demonstrating the viability of the employed tools. The latter yielded four different test-taking types that described distinctive differences between test takers, providing relevant implications for assessment practice.

Background: A priori assumptions about specific behavior in test items can be used to process log data in a rule-based fashion to identify the behavior of interest. In this study, we demonstrate such a top-down approach and created a process indicator to represent what type of information processing (flimsy, breadth-first, satisficing, sampling, laborious) adults exhibit when searching online for information. We examined how often the predefined patterns occurred for a particular task, how consistently they occurred within individuals, and whether they explained task success beyond individual background variables (age, educational attainment, gender) and information processing skills (reading and evaluation skills). Methods: We analyzed the result and log file data of ten countries that participated in the "Programme for the International Assessment of Adult Competencies" (PIAAC). The information processing behaviors were derived for two items that simulated a web search environment. Their explanatory value for task success was investigated with generalized linear mixed models. Results: The results showed item-specific differences in how frequently specific information processing patterns occurred, with a tendency of individuals not to settle on a single behavior across items. The patterns explained task success beyond reading and evaluation skills, with differences across items as to which patterns were most effective for solving a task correctly. The patterns even partially explained age-related differences. Conclusions: Rule-based process indicators have their strengths and weaknesses. Although dependent on the clarity and precision of a predefined rule, they allow for a targeted examination of behaviors of interest and can potentially support educational intervention during a test session. Concerning adults' digital competencies, our study suggests that the effective use of online information is not inherently based on demographic factors but mediated by central skills of lifelong learning and information processing strategies.

Efficiency in reading component skills is crucial for reading comprehension, as efficient subprocesses do not extensively consume limited cognitive resources, making them available for comprehension processes. Cognitive efficiency is typically measured with speeded tests of relatively easy items. Observed responses and response times indicate the latent variables of ability and speed. Interpreting only ability or speed as efficiency may be misleading because there is a within-person dependency between both variables (speed--ability tradeoff [SAT]). Therefore, the present study measures efficiency as ability conditional on speed by controlling speed experimentally with item-level time limits. The proposed timed ability measures of reading component skills are expected to have a clearer interpretation in terms of efficiency and to be better predictors for reading comprehension. To support this claim, this study investigates two component skills, visual word recognition and sentence-level semantic integration (sentence reading), to understand how differences in ability in a timed condition are related to differences in ability and speed in a traditional untimed condition. Moreover, untimed and timed reading component skill measures were used to explain reading comprehension. A German subsample from Programme for International Student Assessment (PISA) 2012 completed the reading component skills tasks with and without item-level time limits and PISA reading tasks. The results showed that timed ability is only moderately related to untimed ability. Furthermore, timed ability measures proved to be stronger predictors of sentence-level and text-level reading comprehension than the corresponding untimed ability and speed measures, although using untimed ability and speed jointly as predictors increased the amount of explained variance.

International large-scale assessments such as PISA or PIAAC have started to provide public or scientific use files for log data; that is, events, event-related attributes and timestamps of test-takers' interactions with the assessment system. Log data and the process indicators derived from it can be used for many purposes. However, the intended uses and interpretations of process indicators require validation, which here means a theoretical and/or empirical justification that inferences about (latent) attributes of the test-taker's work process are valid. This article reviews and synthesizes measurement concepts from various areas, including the standard assessment paradigm, the continuous assessment approach, the evidence-centered design (ECD) framework, and test validation. Based on this synthesis, we address the questions of how to ensure the valid interpretation of process indicators by means of an evidence-centered design of the task situation, and how to empirically challenge the intended interpretation of process indicators by developing and implementing correlational and/or experimental validation strategies. For this purpose, we explicate the process of reasoning from log data to low-level features and process indicators as the outcome of evidence identification. In this process, contextualizing information from log data is essential in order to reduce interpretative ambiguities regarding the derived process indicators. Finally, we show that empirical validation strategies can be adapted from classical approaches investigating the nomothetic span and construct representation. Two worked examples illustrate possible validation strategies for the design phase of measurements and their empirical evaluation.

The study investigates automated and controlled cognitive processes that occur when university students read multiple documents (MDs). We examined data of 401 students dealing with two MD sets in a digital environment. Performance was assessed through several comprehension questions. Recorded log data gave indications about students' time allocation, corroboration, and sourcing. Independent measures were used for reading speed to tap the effects of automatic processing and for working memory and single-text reading comprehension to tap effects of controlled processing, with working memory considered the mental capacity for performing controlled processing. We found that faster readers completed the MD tasks faster and showed more corroboration behavior. At the same time, students skilled in comprehension allocated more time to processing MD tasks and were more likely to show MD-specific behaviors of corroboration and sourcing. Students' success in MD tasks was predicted by reading speed and working memory, with the effect of working memory being mediated by single-text comprehension. Behavioral indicators contributed independently in predicting students' MD comprehension. Results suggest that reading MDs resembles a problem-solving situation where students need to engage in controlled, non-routine processing to build up a comprehensive representation of MDs and benefit from highly automated, lower-level reading processes.

Background: With digital technologies, competence assessments can provide process data, such as mouse clicks with corresponding timestamps, as additional information about the skills and strategies of test takers. However, in order to use variables generated from process data sensibly for educational purposes, their interpretation needs to be validated with regard to their intended meaning. Aims: This study seeks to demonstrate how process data from an assessment of multiple document comprehension can be used to represent sourcing, which summarizes activities for the consideration of the origin and intention of documents. The investigated process variables were created according to theoretical assumptions about sourcing, and systematically tested for differences between persons, units (i.e., documents and items), and properties of the test administration. Sample: The sample included 310 German university students (79.4% female), enrolled in several bachelor's or master's programmes of the social sciences and humanities. Methods: Regarding the hierarchical data structure, the hypotheses were analysed with generalized linear mixed models (GLMM). Results: The results mostly revealed expected differences between individuals and units. However, unexpected effects of the administered order of units and documents were detected. Conclusions: The study demonstrates the theory-informed construction of process variables from log-files and an approach for empirical validation of their interpretation. The results suggest that students apply sourcing for different reasons, but also stress the need of further validation studies and refinements in the operationalization of the indicators investigated.

For many years, reading comprehension in the Programme for International Student Assessment (PISA) was measured via paper-based assessment (PBA). In the 2015 cycle, computer-based assessment (CBA) was introduced, raising the question of whether central equivalence criteria required for a valid interpretation of the results are fulfilled. As an extension of the PISA 2012 main study in Germany, a random subsample of two intact PISA reading clusters, either computerized or paper-based, was assessed using a random group design with an additional within-subject variation. The results are in line with the hypothesis of construct equivalence. That is, the latent cross-mode correlation of PISA reading comprehension was not significantly different from the expected correlation between the two clusters. Significant mode effects on item difficulties were observed for a small number of items only. Interindividual differences found in mode effects were negatively correlated with reading comprehension, but were not predicted by basic computer skills or gender. Further differences between modes were found with respect to the number of missing values.

Introduction: Multiple imputation (MI) is one of the most highly recommended methods for replacing missing values in research data. The scope of this paper is to demonstrate missing data handling in SEM by analyzing two modified data examples from educational psychology, and to give practical recommendations for applied researchers. Method: We provide two examples (N = 589 and N = 621, respectively) based on previous studies of students' self-concepts, mastery goals and performance avoidance goals, and a 7- step tutorial. Then, we produced 20% and 40% missing data under three missing mechanisms by these complete, genuine data sets. The resulting datasets were then analyzed by (1) listwise deletion and structural equation models (SEM), (2) full information maximum likelihood (FIML) with SEM, and (3) MI combined with SEM and pooling. Thus, the results stem from 2 × 3 × 3 conditions. Results: Previous research was replicated by illustrating a practical way to combine MI with SEM and pooling. The assumed factor structure was depicted in both examples with multiply imputed values applied. Discussion: We suggest adding variables to clarify the missing data mechanism, especially for dependent variables as motivation. Such variables might indicate whether missing values in dependent variables are correlated with independent variables (e.g., interest) or the dependent variable itself (e.g. lack of motivation independently of interest).