A Computer Anxiety Model for Elderly Users Interacting with the Web
Thiago Donizetti dos Santos, Vagner Figueredo de Santana · 2019 · Proceedings of the 16th International Web for All Conference (W4A) · doi:10.1145/3315002.3317565
Summary
This paper presents a model for automatically detecting Computer Anxiety (CA) levels in older adults through analysis of their interaction logs while browsing the web. Computer Anxiety — defined as negative emotions and cognition processes evoked during actual or imagined interaction with computer-based technology — is a significant barrier that causes people to avoid or struggle with computers at home, work, and in educational settings. The condition is particularly prevalent among older adults who did not grow up with computers. Traditional methods for measuring CA rely on questionnaires administered by facilitators, which are subjective, cannot scale, and cannot be applied in real-time. The authors conducted a study with 39 elderly participants (ages 61-87, mean 71.92) recruited from a community center for elderly citizens in São Paulo, Brazil, who had never taken computer classes. Participants completed multiple validated questionnaires (Computer Anxiety Rating Scale/CARS, Computer Self-Efficacy/CSE, State-Trait Anxiety Inventory/STAI, Geriatric Depression Scale/GDS, Mini Mental cognitive screening, and System Usability Scale/SUS) and then performed web browsing tasks on a real website (SESC São Paulo) while their interactions were logged by the User Test Logger tool and an eye tracker. Interaction metrics captured included mouse distance, velocity, click counts, pause before click, stroke characteristics, typing speed, and usage graph topology metrics. After excluding participants with depression or cognitive deficit indicators, 31 participants' data were analyzed.
Key findings
Three classification models were built using RandomTree algorithms in Weka, achieving progressively better accuracy as the feature set was refined. Model 1 (all 38 metrics from questionnaires, eye tracker, and interaction logs) achieved 80.64% accuracy. Model 2 (removing eye tracker data) achieved 77.42% accuracy. Model 3 (using only interaction log data — 22 metrics capturable by any web browser) achieved the best accuracy at 83.87%, correctly classifying 26 of 31 participants into no CA, moderate CA, or high CA groups. The counterintuitive result that fewer features yielded better accuracy is explained by the curse of dimensionality with only 31 samples. The interaction-only model (Model 3) identified that people with high CA trigger more total events, click more frequently, take longer to perform tasks, have longer pauses before clicking, and show more task deviations in their usage graphs. Mouse plot visualizations showed strikingly different patterns: high-CA users produced dense, scattered mouse movements with many clicks on non-clickable elements, while no-CA users showed direct, purposeful paths. The study also confirmed the strong inverse relationship between Computer Self-Efficacy and Computer Anxiety — the high-CA group had the lowest CSE scores (48.73 vs 74.57 for no-CA) and lowest SUS ratings (46.14 vs 63.57). Participants who rarely used computers showed the highest CA levels.
Relevance
This research has significant implications for making the web more accessible to older adults by proposing a practical, scalable method for detecting Computer Anxiety in real-time using only standard browser interaction data. Unlike questionnaire-based approaches that require facilitators and cannot operate at scale, Model 3 uses metrics (mouse movements, clicks, pauses, task time) available on any website without special hardware. This opens the possibility of adaptive web interfaces that detect when a user is experiencing anxiety and respond with personalization — simplifying the UI, providing guidance, removing confusing elements, or offering alternative interaction paths. The work broadens the concept of accessibility barriers beyond traditional disability categories to include psychological barriers like anxiety, fear, and low self-efficacy that disproportionately affect older adults. For accessibility practitioners, this is a reminder that usability and accessibility problems do not just create task failures — they can trigger emotional responses that compound the difficulty and lead to technology avoidance. The decision tree models are transparent and interpretable, making them practical for implementation by web developers who can understand and act on the classification logic.
Tags: computer anxiety · older adults · aging · interaction logging · machine learning · usability · accessibility · personalization · user experience · digital inclusion
Standards referenced: WCAG 2.0