Identifying Mixture Components From Large-Scale Keystroke Log Data.

Area of Science:

• Cognitive Psychology
• Educational Technology
• Psychometrics

Background:

• Computer-based writing assessments generate extensive keystroke log data.
• This data offers real-time insights into student writing behaviors.
• Previous research has not fully leveraged this data for cognitive process analysis.

Purpose of the Study:

• To quantify the writing process from a cognitive perspective using keystroke data.
• To develop student writing profiles for personalized instruction.
• To explore the relationship between writing behaviors and cognitive load.

Main Methods:

• Utilized a large dataset of nearly 1,000 student essays from a computer-based writing assessment.
• Applied mixture of lognormal models to analyze pause data from keystroke logs.
• Compared model fit using criteria to determine optimal component numbers (three- vs. four-component models).

Main Results:

• Estimated parameters from pause data were meaningful and interpretable, reflecting varying cognitive processes.
• Findings remained consistent across two different writing genres.
• The mixture model identified distinct patterns in writing processes, including variations in model component preference and mixing proportions related to human-scored quality.

Conclusions:

• Keystroke log data, analyzed with mixture models, can effectively quantify cognitive aspects of the writing process.
• Student writing profiles derived from this analysis can inform writing instruction.
• The study highlights the potential of computational analysis to reveal nuanced writing behaviors and cognitive states.