We resolve a controversy about reading fixations before word-skipping saccades which were reported as longer or shorter than control fixations in earlier studies. Our statistics are based on resampling of matched sets of fixations before skipped and nonskipped words, drawn from a database of 121,321 single fixations contributed by 230 readers of the Potsdam sentence corpus. Matched fixations originated from single-fixation forward-reading patterns and were equated for their positions within words. Fixations before skipped words were shorter before short or high-frequency words and longer before long or low-frequency words in comparison with control fixations. Reasons for inconsistencies in past research and implications for computational models are discussed.
Eye movement control during reading is determined by both perceptual and language processes. Engbert, Longtin, and Kliegl (2002) distinguished three broad categories of models: sequential attention shift (SAS) models, guidance by attentional gradients (GAG) models, and primary oculomotor control (POC) models. These types of models differ in how they weigh the factors for which they are named and in how they conceptualize or implement their interactions (see Reichle, Rayner, & Pollatsek, 2003, for a comparison of the models). All three types of models predict that short, high-frequency, and highly predictable words are skipped more often than are long words, low-frequency words, and words with low predictability, and this prediction is solidly established by experimental research (see Brysbaert & Vitu, 1998, for a meta-analysis). A distinguishing feature of these models, however, concerns their predictions for fixation durations before these skipping saccades.
In principle, there are two different processes that predict increased fixation durations before skipped words: saccade cancelation and parafoveal preprocessing. By definition, in SAS models such as E-Z Reader the next word is the default saccade target. Skipping usually requires a cancelation of the default saccade program and the start of a new one targeting word n + 2 (see, e.g., Engbert & Kliegl, 2001; Reichle, Pollatsek, Fisher, & Rayner, 1998). Therefore, "costs of skipping" were predicted and obtained in simulations (Reichle et al., 2003, Table 1). Moreover, the predicted cancelation costs are likely to decrease with an increase in the frequency of the skipped word because cancelation usually occurs earlier for high-frequency words. Thus, in SAS models longer fixations before skipping are the consequence of saccade cancelation, and cancelation costs should be inversely related to the frequency of the skipped word.
Alternatively, if we allow parallel processing of words within the perceptual span, as GAG models do, longer fixations imply a longer accumulation of parafoveal information. This will increase the probability of complete lexical access and, consequently, of skipping word n + 1 (see, e.g., Engbert et al., 2002; Reilly & Radach, 2003). Therefore, the second reason for skipping might be longer parafoveal processing during the last fixation. Thus, in GAG models long fixations are a cause, not a consequence, of skipping. Note, however, that the link between fixation durations and subsequent skipping is much less tight here than in the case of SAS models, especially for models such as SWIFT (Engbert et al., 2002), where saccade programs are mostly started autonomously.
Finally, POC models have traditionally assumed that saccades are generated from a distribution of saccade amplitudes that may be adjusted to the difficulty of the current text or text segment (McConkie, Kerr, & Dyre, 1994). In the context of skipping, these theories were concerned primarily with effects on saccade amplitudes; they did not predict a strong modulation of fixation durations before skipping, assuming that the next saccade target is determined very early in the fixation, leading to a no-relation hypothesis (Radach & Heller, 2000). …