A Survival Analysis of Fixation Times in Reading

Mattias Nilsson, Joakim NivreCMCL @ ACL 2011 · OpenReview

Mattias Nilsson, Joakim Nivre
Proceedings of the 2nd Workshop on Cognitive Modeling and Computational Linguistics, 2011.

In this paper, researchers explore a novel statistical approach to understanding eye movements during reading. By applying "survival analysis"—a technique traditionally used in medical and biological sciences to model time-to-event data—the authors provide a more nuanced perspective on fixation durations than standard linear models allow.

The Challenge of Modeling Fixation Durations

In reading research, eye movements are characterized by fixations (periods of relative stillness) and saccades (rapid jumps between words). While models for where the eyes move are well-established, modeling when they move remains difficult. Traditional psycholinguistic models often assume that saccade timing is somewhat random or that cognitive processing only influences exceptionally long fixations. Furthermore, fixation durations are typically skewed rather than normally distributed, making standard linear regression techniques problematic.

Nilsson and Nivre address these issues by using a semi-parametric approach to model the entire distribution of fixation times, investigating how linguistic "difficulty" shifts that distribution in real-time.

Methodology: Survival Analysis in Psycholinguistics

The researchers employ two primary survival analysis techniques:

  1. Kaplan-Meier Estimator: A non-parametric method used to establish the baseline "survival" curve of fixations. In this context, "survival" refers to the probability of a fixation continuing past a certain time $t$.
  2. Cox Proportional Hazards Model: This semi-parametric regression models the "hazard rate"—the instantaneous risk that a saccade will occur. This allows researchers to evaluate how various linguistic covariates (e.g., word frequency) multiply this hazard without assuming a specific underlying distribution for the timing.
Kaplan-Meier survival curves for ten readers (subjects a–j), showing the probability of a fixation continuing over time.

To evaluate model performance, the authors utilize the Brier Score to measure prediction error over time and the Concordance Index (C-index) to assess how well the model ranks the durations of pairs of fixations.

Data and Features

The study utilizes the English section of the Dundee Corpus, which contains eye-tracking data from ten readers (subjects a–j) reading newspaper editorials. The models incorporate several linguistic features:

  • Word Frequency: Derived from the British National Corpus.
  • Bigram Probability: Sourced from the Google Web 1T 5-gram corpus.
  • Surprisal and Entropy: Calculated using a probabilistic PCFG parser.

Results and Discussion

The survival curves reveal a critical "Danger Zone" between 150ms and 250ms. As seen in the Kaplan-Meier plots, approximately 80% of fixations are active at 150ms, but this drops to only 20% by 250ms.

Predictive Power

The Cox model, which incorporates linguistic variables, consistently outperformed the baseline Kaplan-Meier model by reducing prediction error. This improvement is most significant during the high-activity 150–250ms window.

Comparison of prediction error (Brier score) over time between the baseline Kaplan-Meier model and the Cox proportional hazards model.

Linguistic Impact and Hazard Ratios

The Cox model provides Hazard Ratios (HR) to quantify the effect of covariates. An HR > 1 indicates an increased risk of a saccade (shorter fixation), while an HR < 1 indicates a decreased risk (longer fixation):

  • Word Frequency: Higher frequency increases the hazard (HR ≈ 1.04), meaning eyes move faster when words are common.
  • Word Length: Longer words significantly decrease the hazard (HR < 1), leading to longer fixations.
  • Surprisal: Higher surprisal generally decreases the hazard, as the eye "survives" longer on unexpected words to facilitate processing.

These results suggest that cognitive processing influences the hazard rate much earlier than some "random-timing" models suggest, constantly modulating the risk of eye movement rather than acting only as a "panic button" for long fixations.

Conclusion and Future Directions

The authors demonstrate that survival analysis is a robust tool for psycholinguistic research, achieving a C-index of approximately 0.60—a meaningful result for complex behavioral data.

Future work could address the slight violations of the proportional hazards assumption (where a covariate's effect might vary over time) using "frailty models" to account for reader variance. Additionally, "competing risks" analysis could be used to differentiate between forward saccades, refixations, and regressions.

Summary updated

BibTeX
@inproceedings{nilsson2011survival,
  author = {Mattias Nilsson and Joakim Nivre},
  title = {{A Survival Analysis of Fixation Times in Reading}},
  booktitle = {CMCL @ ACL},
  year = {2011},
  url = {https://openreview.net/forum?id=e7xrROOME1e},
}
Figure