Validation of a keystroke-level model for a text entry system used by people with disabilities
Heidi H. Koester, Simon P. Levine · 1994 · Proceedings of the First Annual ACM Conference on Assistive Technologies (Assets '94) · doi:10.1145/191028.191061
Summary
This paper develops and validates a keystroke-level model (KLM) to predict how much word prediction software improves text generation rates compared to letter-by-letter typing for users with and without disabilities. The keystroke-level model, originally proposed by Card, Moran, and Newell for mainstream HCI, decomposes user tasks into elementary operations (keystrokes, pointing actions, cognitive decisions) to predict task completion time. The authors adapted this approach specifically for word prediction systems used in assistive technology, where a user types initial letters and the system offers candidate words for selection, potentially reducing the total number of keystrokes needed. Two model variants were tested against actual performance data from both able-bodied and spinal cord injured subjects. Model 1A used parameter values determined independently of subjects' actual performance, while Model 1B derived parameter values from subjects' observed data.
Key findings
Model 1A, using independently estimated user parameters, predicted text generation improvements that differed from actual performance by 11 percentage points for able-bodied subjects but by 53 percentage points for spinal cord injured subjects. This large discrepancy for disabled users highlights that standard user parameter assumptions from able-bodied populations do not generalize well to people with motor impairments — their interaction patterns, timing, and cognitive-motor trade-offs differ substantially. Model 1B, which used parameter values derived from each subject's actual data, achieved much better accuracy with an average error of only 6 percentage points across all subjects. This demonstrates that the keystroke-level modeling approach is fundamentally sound for predicting word prediction benefits, but requires calibration with individual user data to be accurate for people with disabilities. The finding that word prediction benefits vary significantly between users underscores the importance of individualized assessment when prescribing assistive text entry tools.
Relevance
This paper makes an important methodological contribution to assistive technology evaluation by applying formal HCI performance modeling to predict the benefits of word prediction for users with disabilities. The finding that generic user parameters fail for spinal cord injured users but individualized parameters succeed has broad implications: assistive technology effectiveness cannot be reliably predicted from able-bodied performance data, and evaluation methods must account for the diversity of motor abilities among users with disabilities. This remains a relevant challenge today as word prediction and autocomplete features are ubiquitous in mobile and desktop computing. For practitioners, the paper provides a principled framework for understanding when word prediction will actually help a given user — it is not universally beneficial, and the cognitive overhead of scanning prediction lists can sometimes offset keystroke savings, particularly for faster typists.
Tags: text entry · word prediction · keystroke-level model · spinal cord injury · usability · human-computer interaction · performance modeling · assistive technology