How important are linguistic factors in word skipping during reading?

Abstract

The probability of skipping a word is influenced by its processing ease. For instance, a word that is predictable from the preceding context is skipped more often than an unpredictable word. A meta-analysis of studies examining this predictability effect reported effect sizes ranging from 0 to 13%, with an average of 8%. One study does not fit within this picture and reported 23% more skipping of Dutch pronouns in sentences in which the pronoun had no disambiguating value (e.g. 'Mary was envious of Helen because she never looked so good') than in sentences where it did have a disambiguating value (e.g. 'Mary was envious of Albert because she never looked so good'). We re-examined this ambiguity in Dutch using a task that more closely resembles normal reading and observed only a 9% difference in skipping of the pronoun, bringing this linguistic effect in line with the other findings.

Full text

1
How Important are Linguistic Factors in Word Skipping during Reading?
Denis Drieghe1, Timothy Desmet1, and Marc Brysbaert2
1Ghent University, Belgium
2Royal Holloway, University of London, UK
Address: Denis Drieghe
Department of Psychology
Tobin Hall
University of Massachusetts
Amherst
MA 01003
USA
Tel. ++ 1 – 413 – 545 – 6662
e-mail: denis.drieghe@UGent.be
Running Head: Word skipping during reading

2
Acknowledgments
The first two authors contributed equally to this work. Both Denis Drieghe and
Timothy Desmet are postdoctoral fellows of the Fund for Scientific Research
(Flanders, Belgium). We thank Barbara Juhasz, Simon Liversedge and Françoise Vitu
for many helpful comments on the previous draft of this article.

3
Abstract
The probability of skipping a word is influenced by its processing ease. For instance,
a word that is predictable from the preceding context is skipped more often than an
unpredictable word. A meta-analysis of studies examining this predictability effect
reported effect sizes ranging from 0% to 13%, with an average of 8% (Brysbaert,
Drieghe, & Vitu, 2005). One study does not fit within this picture: Vonk (1984)
reported 23% more skipping of Dutch pronouns in sentences where the pronoun had
no disambiguating value (e.g. “Mary was envious of Helen because she never looked
so good”) than in sentences where it did have a disambiguating value (e.g. “Mary was
envious of Albert because she never looked so good”). We re-examined this
ambiguity in Dutch using a task that more closely resembles normal reading and
observed only 9% difference in skipping of the pronoun, bringing this linguistic effect
in line with the other findings.

4
Introduction
In order to understand what determines eye guidance in text reading, many studies
have focused on visual and linguistic factors to explain eye movement variables such
as fixation times and saccade lengths. Whereas these investigations have led to a huge
increase in our understanding of eye movements in reading (Liversedge & Findlay,
2000; Rayner, 1978, 1998), a number of issues remain disputed (Starr & Rayner,
2001). One of these controversies deals with the question of what precisely influences
the eye guidance system when it decides not to make a saccade to the next word but to
fixate the following word (i.e. word skipping). In this article we will focus on the
relative contributions of visual versus linguistic factors in this decision, an issue we
will show to be at the core of the ongoing debate concerning the skipping of words in
reading.
Although eye movement patterns during reading seem almost incomprehensibly
complex, a lot can be gained from considering every eye movement as an individual
decision of where and when to move the eyes. Interestingly, there is a big difference
between the variables that influence these two decisions. The decision of when to
move the eyes away from a word (reflected in fixation times on that word) is
primarily determined by the processing ease of the word. A very robust finding is that
readers will spend more time looking at a low-frequency word than at a high-
frequency word (e.g. Inhoff & Rayner, 1986; Rayner & Duffy, 1986; Schilling,
Rayner & Chumbley, 1998). Likewise, a word that is predictable from the preceding
context will be looked at for a shorter time than a neutral word (e.g. Balota, Pollatsek,
& Rayner, 1985; Binder, Pollatsek, & Rayner, 1999; Ehrlich & Rayner, 1981; Rayner
& Well, 1996). Even though visual factors can influence the gaze duration on a word,
the linguistic properties of the word account for quite a large part of the variance in
fixation times. On the other hand, the decision of where to move the eyes seems to be
mostly determined by visual factors, such as the length of the currently fixated word
and the lengths of the next words. For instance, Rayner (1979) showed that readers
have a tendency to make the first fixation on a word slightly left of the center of that
word (but see White & Liversedge, 2004 for an example of a linguistic influence on
landing site).

5
A first indication of the disputed nature of the word skipping phenomenon is that it is
difficult to place within this commonly used when/where dichotomy. Robust
influences of both a low-level visual nature and a high-level linguistic nature have
been shown to affect skipping behavior. A typical low-level visual effect on word
skipping is the effect of launch site: The closer the eyes are to the parafoveal word,
the higher the probability that this word will be skipped on the next saccade (Kerr,
1992; Rayner, Sereno & Raney, 1996; Vitu, O’Regan, Inhoff & Topolski, 1995). The
most robust empirical finding in word skipping related to the characteristics of the
word itself is the effect of word length: Readers tend to skip short words more often
than long words (e.g. Rayner, 1979; Rayner & McConkie, 1976; Vitu et al., 1995)1.
This could be due either to the length of the word, or to the processing ease of the
word (many short words are high-frequency words or syntactic function words).
Processing ease has also been shown to influence skipping behavior when word
length is controlled for: A word that is predictable from the preceding context is
skipped more often than a word that is not predictable (e.g. Balota, et al., 1985;
Drieghe, Brysbaert, Desmet, & De Baecke, 2004; Ehrlich & Rayner, 1981; Rayner,
Binder, Ashby & Pollatsek, 2001; Rayner & Well, 1996) and a high-frequency word
is skipped more often than a low-frequency word (e.g. Henderson & Ferreira, 1993;
Radach & Kempe, 1993; Rayner & Fisher, 1996; Rayner et al., 1996). The
predictability and frequency effects suggest that a skipped word has already been
processed to a certain degree while the eyes were fixating on the previous word. Quite
some debate in the eye movement literature focuses precisely on the extent to which a
word can be processed in parafoveal vision and what effects this has on eye
movement control (e.g. Radach & Kennedy, 2004). Whereas some models posit that
the eye movement system can skip a word only when the word has been completely
recognized on the prior fixation or when full recognition is imminent (e.g. the E-Z
Reader model, Reichle, Rayner, & Pollatsek, 2003), other models claim that word
skipping is based on coarser information and that it entails an educated guess taking
into account factors such as word length and only very partial word identification. In
1 This observation has also been made when readers are asked to “fake” eye movements while scanning
through z-strings. Based on the similarities between the skipping patterns of string scanning and normal
reading Vitu et al. (1995) concluded that predetermined oculomotor strategies are an important
determinant of eye movement control in normal reading. However, these conclusions have been
questioned by subsequent research (Rayner & Fischer, 1996) and experimental manipulations have
been shown to affect skipping of z-strings and normal words in different ways (Drieghe, Brysbaert &
Desmet, 2005).

6
the EOVP model (Brysbaert & Vitu, 1998) for instance, the main determiners of the
decision to skip a word are the length of the word and the experience the system has
built up with how often a word of a certain length at a certain distance can be skipped
without hindering overall text comprehension. This decision can be made with very
limited information about the identity of the word. Other models, such as the SWIFT
model (Engbert, Longtin & Kliegl, 2002; Engbert, Nuthmann, Richter & Kliegl,
2005) or the Glenmore model (Reilly & Radach, 2003; 2006) can be placed
somewhere between the 2 views mentioned above in terms of how much parafoveal
processing they assume to take place prior to skipping. In the SWIFT model, for
instance, saccades will be oriented towards words that have the highest level of
excitation, which occurs at intermediate steps of lexical processing. As a
consequence, the more processing of the parafoveal word (word n + 1) has occurred,
the higher the chances that the level of excitation for this word will already have
passed its peak and is surpassed by the level of excitation of the subsequent word
(word n + 2). Word n + 2 will then win the competition of becoming the target for the
next saccade and word n + 1 will be skipped. As a consequence, SWIFT allows a
word to be skipped even when the level of lexical processing of word n + 1 has not
yet reached the amount assumed by the E-Z Reader model (e.g. a word n + 2 with a
very high level of excitation would increase the chances of the word n + 1 being
skipped, regardless of the level of excitation associated with word n + 1). On the other
hand, SWIFT usually assumes a lot more processing of word n + 1 before it can be
skipped than is assumed in the EOVP model.
It is important to stress that most studies that find linguistic influences on word
skipping did control for visual factors and vice versa. Hence, any comprehensive
model on word skipping will have to take into account both visual and linguistic
factors. In the E-Z Reader model the effect of word length on skipping is incorporated
because the model assumes an inverse relation between the extraction of letter
information and the distance of a letter from the center of the visual field. So the
further away the eyes are from the target word, the more time will be needed to
complete the first phase of word recognition. Because word skipping depends on word
recognition, this slowdown in the word recognition process leads to a smaller chance
that the word will be skipped. Basically the same principle features in the SWIFT
model: Due to a reduced efficiency of information extraction of a word at a large

7
eccentricity, chances are lower that this word will have passed its peak in excitation.
Whereas the EOVP model states that the initial decision whether or not to skip the
next word will be based on word length and the distance of the word from the current
fixation location, it does allow some limited saccade target adaptation by incoming
linguistic information after this initial decision to skip or not has been made.
A convincing illustration of the importance of incorporating both sources of influence
on word skipping is provided by Drieghe, Rayner and Pollatsek (2005). In their first
experiment they showed that a high amount of lexical processing can influence the
decision of the eye movement system to skip a word. In a sentence such as “The
doctor told Fred that his drinking would damage his liver very quickly.” the word
“liver” is highly predictable from the preceding context. The skipping rates of this
predictable target word were compared to sentences in which the target word was
replaced either by a neutral word (e.g. heart) or by an illegal non-word that was
visually identical to the predictable word with the exception of a single letter (e.g.
livor). Drieghe et al. showed that the predictable word was skipped more often than
the neutral word, replicating the findings by Balota et al. (1985). However, no
difference was reported between the skipping rates of the neutral word (“heart”) and
those of the visually similar non-word (“livor”). In other words, only if the word
completely matched the predictable word did an effect of predictability on skipping
rate emerge. However, in a second experiment, evidence was found that
orthographically illegal non-words (e.g. hxt) were skipped too often to be attributed to
factors such as saccadic error (at a close launch site 31 % of the orthographically
illegal non-words were skipped)2. If the decision to skip was exclusively determined
by the success of recognizing the word in the parafovea, non-words would not be
skipped so often. In all likelihood, factors that are more low-level than the lexical
processing of the parafoveal word (such as the short length of these orthographically
illegal non-words) were playing an important role in causing the eyes to skip these
non-words. So, from these two experiments it seems that both visual and linguistic
factors play an undeniable role in determining skipping behavior. It becomes clear
that the decision of which theoretical model gives the best account on how the system
2 Targeting errors are far from rare. A recent analysis by Nuthmann, Engbert, and Kliegl (2005)
suggests that more than 10% of saccades miss their intended target.

8
arrives at skipping words, will be answered on the basis of the relative importance of
these factors instead of choosing one type of factor over the other.
In an attempt to get a more accurate view of the relative importance of visual versus
linguistic influences on word skipping, Brysbaert, Drieghe, and Vitu (2005; for an
earlier version see Brysbaert & Vitu, 1998) carried out a meta-analysis of all word-
skipping studies that manipulated the processing ease of words and that reported word
lengths. A distinction was made between a first group of studies that manipulated the
processing ease of the target word in terms of word characteristics (e.g. the frequency
of the word) and a second group of studies that manipulated processing ease in terms
of contextual predictability. For the first group of studies Brysbaert et al. reported a
consistent 5% difference in skipping rate between the easy and the difficult words,
whereas the word length effect ranged from a negligible 2% skipping rate for nine-
letter words to over 50% skipping rates for two-letter target words. For the second
group of studies the average effect of contextual predictability on word skipping
amounted to an 8% difference between the predictable target words and the neutral
words. The effects of word length were identical to those found for the first group of
studies. From these findings the authors concluded that to predict the skipping rate of
a word, it was more useful to know how long it is than how easy it is to process.
Given the theoretical importance of these data, it is interesting to look at the range of
effects of linguistic factors on word skipping. Whereas the meta-analysis revealed
very similar ranges of processing ease effects in the studies that manipulated word
characteristics (effects of 0% to 13%) and the studies that manipulated context
predictability (effects of 0% to 12%), there was one major exception in the latter
group of studies. Vonk (1984) reported a difference of 23 % in skipping rate between
the predictable and neutral condition; almost twice the effect of the largest skipping
rate of 12% and almost three times as much as the average skipping rate of 8%. This
is the only study that reports a linguistic effect that appears to be comparable in size
with the visual word-length effects. Therefore it is crucial for our understanding of the
role of linguistic and visual factors in word skipping (and eye movement behavior in
general) to have a closer look at this experiment.
Vonk (1984) compared the skipping rates of Dutch pronouns in sentences such as (1).

9
(1a) Alex beloog Andy omdat hij onraad rook.
[Alex lied to Andy because he smelled trouble.]
(1b) Alex beloog Anna omdat hij onraad rook.
[Alex lied to Anna because he smelled trouble.]
Vonk investigated whether readers are able to process a sentence in a selective way;
i.e. whether they pay more attention to words if they contain new information than if
their meaning is made redundant by the preceding clause. In sentence (1a), where two
masculine names are used, the pronoun “he” does not provide disambiguation
information concerning the intended antecedent of the pronoun. In sentence (1b),
where a masculine and feminine name are used, the gender of the pronoun is
informative as to which entity in the previous clause is the correct antecedent of the
pronoun. Vonk reported a skipping rate of 40% of the pronoun in sentence (1a) and
17% in sentence (1b). She interpreted these skipping rates as evidence for a rational
selection of information in sentence reading. Given that the pronoun did not offer any
extra information in terms of the correct antecedent of the pronoun in sentence (1a), it
was skipped more often than in sentence (1b), where the pronoun was informative.
However, whereas most studies use a normal reading task to investigate skipping
behaviour, Vonk (1984) used a referent naming task, in which participants had to
vocalize the intended antecedent of the pronoun. So, participants pressed a button in
order to let a sentence appear on the screen. Their task was to name the correct
referent for the pronoun as fast as possible. This means that participant could name
the antecedent of the pronoun before they had read the entire sentence. After the
participant named the referent of the pronoun, he/she had to press the button again to
end the trial. It is not inconceivable that participants quite quickly grasped that a
combination of two antecedents sharing the same gender made the identity of the
pronoun referentially ambiguous and therefore useless for the task at hand (naming
the intended antecedent). On the other hand, when one of the names is masculine and
one is feminine, it is easy to see that simply looking at the pronoun would
immediately provide you with the right response to the task (leading to almost no
skipping in that case). Consequently, it is very likely that the task that Vonk used,
exaggerated the difference between fixating ambiguous and unambiguous pronouns.

10
Therefore, and especially given the rather special position this study holds in reported
effects of linguistic factors on word skipping, it would be worthwhile to examine
whether this effect would occur in the same size in an experiment using a task that
more closely resembles normal reading. If the size of the pronoun ambiguity effect in
“normal” reading would be reduced to the 0% - 12% range – the range reported in all
the other word skipping studies looking at linguistic variables – this would be
important for the discussion on the relative importance of visual versus linguistic
factors on word skipping. Finding out the actual size of the effect would also be
important for people who try to model skipping data, as the size of the effect reported
by Vonk (1984) clearly constitutes a statistical outlier. If on the other hand, the size of
the effect upholds under normal reading conditions, this would shed new light on how
pronoun resolution can effect word skipping. Like the original experiment, the current
experiment was conducted in Dutch. This language has the extra advantage for word
skipping research in that the male pronoun hij (he), has the same number of letters as
the female pronoun zij (she), which permits perfect matching on word length.
EXPERIMENT
Participants. Thirty-six members of the Ghent University community participated in
this experiment. All participants were native speakers of Dutch and had normal or
corrected-to-normal vision. They were paid 10€ for their participation.
Apparatus. Eye movements were recorded by a Senso-Motoric Instruments (SMI
Eyelink) video-based pupil tracking system. Viewing was binocular but eye
movements were recorded from the right eye only. A high speed video camera was
used for recording. It was positioned underneath the monitored eye and held in place
by head-mounted gear. The system has a visual resolution of 20 seconds of arc.
Fixation locations were sampled every 4 ms and these raw data were used to
determine the different measures of oculomotor activity during reading. The display
was 69 cm from the subject’s eye and three characters equalled 1° of visual angle. A
chin rest was used to reduce head movements during the experiment.

11
Materials. Thirty-two target sentences were created3. Each target sentence had two
versions based on the disambiguating value of the pronoun: the gender of the pronoun
was either ambiguous (both antecedents had the same gender as the pronoun) or
unambiguous (the two antecedents had a different gender and thus the gender of only
one antecedent corresponded to the gender of the pronoun). Half of the items used the
masculine pronoun (“hij” [he]), whereas the other half of the items used the feminine
pronoun (“zij” [she]). We included both the masculine pronoun “hij” and the feminine
pronoun “zij”, because using only one of them could lead to participants developing a
strategy in the unambiguous cases where as soon as you see the two names you know
which one is the correct antecedent (independent of the information carried by the
pronoun). An example of the 2 conditions, both with a male and a female pronoun, is
given in Table 1 together with their translations in English.
INSERT TABLE 1 ABOUT HERE
It is important to note that the type of material that was used by Vonk (1984),
has subsequently mainly been used for the study of the role of implicit causality on
pronoun resolution. The implicit causality bias is a semantic characteristic of verbs
such that event descriptions (based on these verbs) sometimes implicitly indicate the
cause of the event (e.g. Garnham, Oakhill, Cruttenden, 1992; Garnham, Traxler,
Oakhill, & Gernsbacher, 1996; Greene & McKoon, 1995; Long & De Ley, 2000;
McDonald & MacWhinney, 1995; McKoon, Greene, & Ratcliff, 1993). For instance,
when participants are presented with the sentence "John amazed Mary" and are asked
to indicate the cause of this event, they will predominantly attribute the cause to the
agent of the event (i.e. the first-mentioned name in active sentences, such as "John").
However, when participants are presented with the sentence "John admired Mary",
they will predominantly attribute the cause to the patient of the event (i.e. the second-
mentioned name in active sentences, such as "Mary"). For this reason, verbs such as
"amaze" are called N1 verbs and verbs such as "admire" are called N2 verbs. The
implicit causality bias also has an online effect in that sentences which contain an
explicit cause that is congruent with the implicit bias (e.g. "John admired Mary
because she was a very smart woman") are read faster than sentences which contain
3 All materials are available from the first author upon request, denis.drieghe@ugent.be

12
an explicit cause that is incongruent with the implicit bias (e.g. "John admired Mary
because he liked very smart women") (Stewart, Pickering, Sanford, 1998). In order to
avoid any confounding effects of this much-studied implicit causality bias in our
study, we made sure to have as many N1-biased verbs as N2-biased verbs in our
stimuli and that the explicit cause in our items was always congruent with the implicit
causality bias.
Procedure. Before the experiment started, participants were informed that the study
was about the comprehension of sentences that were displayed on a computer screen.
Sentence administration was self-paced: Participants started and stopped sentence
presentation by pressing on a button. Each sentence was presented as a whole.
Participants were asked to read at their normal speed, and to answer any
comprehension question that would follow the passage. Questions, which were simple
true/false statements, followed on one fourth of the trials. The participants had no
difficulty answering these questions (the overall question answering accuracy rate was
96 %). The initial calibration of the eye-tracking system generally took approximately
10 min and consisted of a standard nine-point grid. Following the initial calibration
the participant was given 10 practice trials to become familiar with the procedure
before reading the experimental sentences. The 32 experimental sentences were
embedded in a pseudo-random order in 110 filler texts. These filler texts were part of
other reading experiments conducted at our lab. Each participant was presented with
one of the two possible variants of the critical text fragments according to a Latin
square design. Participants completed a single session lasting about 50 min,
containing 142 text fragments to read.
RESULTS
Our primary interest in this experiment was the probability of skipping the pronoun
during the first pass through the text. In addition to the skipping probability, we also
examined the fixation times on the pronoun. Due to the fact that the pronoun was a
three-letter word, it is not surprising that when the pronoun was fixated it received
only a single fixation in 98% of the cases. Therefore we will restrict the fixation
duration analyses on the pronoun to those occasions when there was a single fixation.
Besides fixation times on the pronoun we will also examine the first-pass time of the

13
region after the pronoun, which was defined as the two words following the pronoun,
to see whether the effect of gender cue resulted in any differences in eye movements
after the pronoun was or was not skipped. The first-pass time is the sum of all the
fixations on a region prior to an eye movement to another region in the text. The
overall reading times will also be computed in order to establish whether one type of
sentences was more difficult to read than the other. Only 1.8 % of the data were
removed from the analyses because of track loss or because the fixation was shorter
than 100 ms (see Morrison, 1984; Rayner, Sereno, Morris, Schmauder, & Clifton,
1989, for justification). All analyses were run over participants (F1-analyses) and
items (F2-analyses).
Skipping the pronoun. The skipping probabilities of the pronoun are shown in Table 2.
We ran a repeated-measures ANOVA on the skipping probabilities of the pronoun
with gender cue (presence versus absence of a gender cue) and the gender of the
pronoun (“he” or “she”) as independent variables4. The 5% effect of gender cue was
not significant [F1(1,35) = 2.14, p > .10; F2(1,15) < 1, n.s.] but the effect of pronoun
gender was significant [F1(1,35) = 7.90; p < .01; F2 (1,15) = 9.68; p < .01]: On
average the male pronoun (“hij”, he) was skipped 10 % more often than the female
pronoun (“zij”, she). There was no significant interaction between these two factors
[F1(1,35) = 2.10, p > .10; F2(1,15) = 2.77, p > .10]. However, to make sure that we
did not miss any effect of the gender cue we analyzed the skipping probabilities of the
“he” and “she” conditions separately. These planned comparisons showed that the
pronoun “he” was skipped 9 % more often when there was no gender cue than when
there was one [t1(35) = 2.07, p < .05; t2(35) = 2.77, p = .052]. The same effect was
not present in the “she” conditions (all t’s < 1).
INSERT TABLE 2 ABOUT HERE
Single fixation times on the pronoun. As can be seen from the fixation times shown in
Table 2 there was no effect of gender cue (all F’s < 1) and no effect of the pronoun
4 Since we made sure that the antecedent of the pronoun was always congruent with the N1 or N2
implicit causality bias of the verb, this factor did not cause any significant effects throughout all the
analyses reported here (all F’s < 1). Therefore this factor was not taken into account while reporting the
other analyses.

14
gender on the single fixation times on the pronoun (all F’s < 1). The interaction
between these variables was not significant either (all F’s < 1).
First-pass time on the region after the target. The effect of gender cue was not
significant on the first-pass time of the region after the target [F1(1,35) = 2.58, p >
.10; F2(1,15) = 1.63, p > .20]. On the other hand, the effect of pronoun gender was
significant in the participants analysis [F1(1,35) = 32.08, p < .001] but was not in the
items analysis [F2(1,15) = 2.79, p > .10]. So there was some indication that the region
following the target word was read faster in the “he” conditions than in the “she”
conditions. The interaction between these 2 factors was not significant (all F’s < 1).
Total reading times of the sentences. The effect of gender cue was not significant in
the analyses of the total reading times (all F’s < 1) but there was a significant effect of
pronoun gender, although again restricted to the participants analysis [F1(1,35) =
7.47, p < .01; F2(1,15) = 1.44, p > .20]. The sentences containing a “he” pronoun
appear to be read on average 104 ms faster than the sentences containing a “she”
pronoun. The interaction between these two factors was not significant (all F’s < 1).
Fixation prior to the target. Given the rather surprising finding that our results were
different when the pronoun was masculine compared to when it was feminine, we
conducted an analysis on the fixation duration prior to either the landing on or the
skipping of the pronoun. Although unlikely, should any difference between the
stimulus sets containing the masculine or the feminine pronoun contribute to our
skipping data, these differences would most certainly be apparent in these analyses.
These analyses could also be informative if any of our manipulations had an effect
prior to our region of interest, the pronoun. Table 3 reports both the mean fixation
duration and the launch site as a function of whether the next saccade either lands on
the pronoun or skips it.
As is apparent from the means, the analysis of the fixation duration prior to skipping
showed no effects of gender cue, pronoun gender or an interaction between these 2
variables (all F’s < 1). The same pattern was observed for the fixation duration prior
to landing (all F’s < 1). For the analysis of the launch site prior to the skipping of the
pronoun there was no effect of gender cue (F1(1,28) = 1.67, p > .20; F2(1,15) = 1.93,

15
p > .10) and no effect of pronoun gender (all F’s < 1). The interaction between these
two factors was also not significant (all F’s < 1). The analysis of the fixation duration
prior to landing did also not result in any significant differences: there was no effect
of gender cue (all F’s < 1), no effect of pronoun gender (F1(1,31) = 2.48, p > .10; F2
< 1, n.s.) and the interaction was also not significant (F1(1,31) < 1, n.s.; F2(1,15) =
2.40, p > .10). Collapsing over all the data, we did replicate the well-documented
effect of launch-site: the eyes were 4 character positions closer to the pronoun when
they skipped the pronoun than when they did not (resp. 3.7 character positions versus
7.7 character positions, t1(35) = -14.74, p < .001; t2(31) = -11.86, p < .001). The
fixation duration prior to skipping was also 15 ms shorter than prior to landing on the
pronoun, although this effect was only marginally significant in the items analysis
(resp. 223 ms versus 238 ms, t1(34) = 3.44, p < .01; t2(31) = 1.92, p = .06)5.
INSERT TABLE 3 ABOUT HERE
Discussion
The main goal of the current study was to test whether the effect size of a linguistic
factor (i.e. a gender cue) on the skipping rate of pronouns, as reported by Vonk
(1984), could be replicated in a task that more closely resembles normal reading.
Vonk reported that a gender-marked pronoun was skipped 23 % more often when the
two potential antecedents had the same gender than when the two antecedents were of
a different gender. The size of this effect is quite unusual, given that a recent meta-
analysis (Brysbaert et al., 2005) showed that all other reported effects of a linguistic
origin on word skipping were in the range of 0% - 13%. Therefore, it was important to
re-examine this study as it clearly constitutes an outlier in terms of linguistic
influences on word skipping. If the size of the effects in normal reading are of a
5 Looking at the launch sites analyses, it is apparent that the saccades resulting in the skipping of the
pronoun more often came from the word prior to the pronoun (i.e. ‘omdat’ [because]) than was the case
when the saccade resulted in landing on the pronoun. In all likelihood the reduced fixation duration
prior to skipping reflects the high frequency of the word omdat [because]. Recently, there has much
discussion in the literature on whether or not the fixation duration prior to skipping is inflated (e.g.
Drieghe et al., 2005). This observation is important because it is quite often seen as one the ways to
distinguish between the different models of eye movement control in reading (for a discussion see
Kliegl & Engbert, 2005). However, for the present study we do not have the amount of data necessary
to conduct an analysis of the fixation duration prior to skipping/landing that is sufficiently controlled
for launch site. Hence we will not go further into this issue.

16
comparable magnitude to those observed in the original study, this would show a
stronger influence of linguistic factors on the decision to skip a word than previously
observed. It would mean that at least some linguistic factors (pronoun ambiguity) can
strongly influence skipping behaviour. It would also mean that word skipping is very
instrumental in examining early effects of pronoun resolution, because one of the
major debates in the psycholinguistic literature on implicit causality bias is exactly
about how early this bias can influence pronoun resolution (e.g. Stewart et al., 2000).
However, there were a priori reasons to doubt whether we would replicate such a
large effect in normal reading: the task in the Vonk experiment was to vocalize as fast
as possible the name of the antecedent of the pronoun. It is not inconceivable that this
task elicited different processing strategies (focusing more on the gender information
of the pronoun) than those used in other reading experiments investigating the effect
of linguistic variables on word skipping. In the current study the task was similar to
other word skipping studies, namely reading for comprehension.
Our results clearly show that the large effects of gender cue on the skipping
probability of the pronoun reported by Vonk (1984) were a consequence of the task
used. Using the same type of materials, we found a non-significant effect of 5% more
skipping of the pronoun when it was uninformative than when it was informative. At
the same time, there seemed to be an unexpected difference between the masculine
pronoun hij and the feminine pronoun zij. The pronoun hij was skipped 10% more
often than the pronoun zij and showed the clearest evidence for an effect of gender
cue (9% vs. 1% for zij). Moreover, the latter effect was statistically significant,
replicating the findings reported by Vonk (1984), but bringing them within the range
of effect sizes that was found with other linguistic variables. Also from the reading
times it was clear that sentences with the pronoun hij were easier to understand than
sentences with the pronoun zij.
The fact that we replicated the effect due to the pronoun informativeness, even though
in a much less dramatic way than in the original Vonk (1984) study, indicates that
readers sometimes look at parts of a sentence in a selective way. In this experiment
they preferred to move the eyes further into the text if the gender of the antecedents
made the pronoun redundant. Note that this is very likely an effect of strategically
positioning the eyes and is not solely due to the processing ease of the pronoun itself.

17
The fact that no effects whatsoever were observed in the fixation times of the pronoun
as a function of our manipulations shows that the skipping rates in this experiment do
not merely reflect the processing ease of the pronoun (although this lack of effects
could also originate from a floor effect given the very short fixation times). When a
word is redundant, readers skip the word more often so that they can position the eyes
further in the text. Before dealing with the difference we observed as a function of the
gender of the pronoun, we will first discuss the implications of the removal of the
Vonk (1984) data from the charts of effects on word skipping in normal reading.
Theoretical accounts of the word skipping phenomenon differ greatly in the extent to
which they allow parafoveal preprocessing to determine the target of a between-word
saccade. Models such as the E-Z Reader model (Reichle, Rayner, & Pollatsek, 2003)
state that in order for a word to be skipped the system either must have recognized the
word on the prior fixation or that full recognition must be imminent. Other models,
such as the EOVP model (Brysbaert & Vitu, 1998) claim that word skipping is
primarily based on coarse information concerning the parafoveal word, such as its
length and its distance from the current fixation position. Only in a second stage can
parafoveal word processing affect the initial decision, either by canceling an intended
saccade to the next word or by canceling the initial decision to skip the next word.
Without going into the details of these models, it is important to note that all current
models that make explicit claims on word skipping in reading incorporate influences
of both a linguistic and a visual nature. So the main argument is about which of the
influences on word skipping is the main player in determining the decision whether or
not to skip the following word. If low-level visual factors (e.g. word length) are the
main players, this would be a strong indication that the (initial) decision to skip a
word is primarily based on relatively coarse information concerning that word. If on
the other hand, linguistic factors are the most important factor in determining skipping
behavior, this would clearly indicate that it is the advanced state of recognizing the
parafoveal word that triggers the word skipping saccade.
Our present findings constitute an important addition to the meta-analysis reported by
Brysbaert, et al. (2005), which looked at the relative importance of visual versus
linguistic factors in determining word skipping behavior. This meta-analysis showed
that in order to predict the skipping rate of a word it was more informative to know

18
the word length than to know how easy the word was to process. One study did not fit
in that picture: Vonk (1984) reported an effect of a linguistic origin that was
comparable in size to the reported word length effects. By removing this study from
the charts we add further evidence to the claim that linguistic influences are not the
main determinant of the decision to skip the upcoming word. Although word skipping
is affected by the informativeness and the difficulty of the word, it is more
informative to know the length of the parafoveal word in order to predict whether it
will be skipped or not.
Our results also suggested that the effect of gender cue was restricted to the skipping
of the male pronoun hij (he). This finding was quite surprising since no such effects
were observed in the Vonk (1984) study or anywhere else in the literature. However,
the presence of another unexpected finding, namely that the male pronoun hij (he)
was overall skipped more frequently than the female pronoun zij (she) can point us in
the direction of a plausible explanation. Note that we had included both pronouns
simply to exclude the possibility that always the female (if we would have used only
zij) or male name (if we would have used only hij) would be the correct antecedent.
Within a trial we always compared skipping of the same pronoun in a condition where
both names had a different gender with a condition where both names had the same
gender. We did not expect that this factor would have an influence on skipping,
because both pronouns have the same length and because they are both extremely
frequent (the pronoun hij [he] has a a log frequency per million words of 4.05, the
pronoun zij [she] has a log frequency per million words of 3.42, see Baayen,
Piepenbrock, & Van Rijn, 1993). However, a peculiarity in Dutch is that the female
singular pronoun zij (she) is identical to the plural pronoun zij (they). In other words
when readers encounter a sentence such as “Laura verontschuldigde zich bij Simon
omdat zij... [Laura apologized to Simon because she…]” it is perfectly legal to
continue the sentence in such a manner that zij refers to both Laura and Simon (e.g.
Laura verontschuldigde zich bij Simon omdat zij vroeger afgesproken hadden [Laura
apologized to Simon because they arranged to meet earlier]. Due to this peculiarity in
Dutch it is not inconceivable that readers, upon encountering zij, process both the
singular female and the plural possibility. If this were the case, we would expect to
observe in the current experiment both lower skipping rates for the pronoun zij (she)
and somewhat longer total reading times in the zij (she) sentences, indicating the

19
higher processing costs associated with processing both the female singular and the
plural option. We observe indications of both these predictions in the current
experiment (although one has to be very careful with interpreting the total reading
times, given the fact that pronoun gender was manipulated across 2 halves of the
stimulus set). This phenomenon is clearly beyond the scope of the present study and a
post-hoc interpretation that should be properly tested by matching the sentences used
for the pronouns. However, we do see this explanation as the most plausible cause
both for the lower skipping rates we observed in the female pronoun conditions and
for the absence of the gender cue effect in those conditions: the fact that “zij” could
refer to both a female singular antecedent and plural antecedents would in all
likelihood destroy any effect of redundancy we could observe in these conditions.
To summarize, in this study we re-examined a prior experiment (Vonk, 1984) that was
the only one to report an effect of a linguistic nature on word skipping rates, that was
in size comparable to the effects reported for word length. Using a task more closely
resembling normal reading, we were able to show that the size of this effect was well
within the range of previously reported effect sizes of linguistic factors in word
skipping. The predictability effect was 5% if we looked at all the sentences of the type
used by Vonk, and 9% if we only looked at the sentences for which there was no
discussion about the distinction between informative and uninformative pronouns.
This finding is further evidence for the statement that in order to predict whether a
word will be skipped or not, it is more instrumental to know the length of the word
than its processing ease.

20
References
Balota, D. A., Pollatsek, A., & Rayner, K. (1985). The interaction of contextual
constraints and parafoveal visual information in reading. Cognitive
Psychology, 17, 364-390.
Binder, K. S., Pollatsek, A., & Rayner, K. (1999). Extraction of information to the left
of the fixated word in reading. Journal of Experimental Psychology: Human
Perception and Performance, 25, 1162 – 1172.
Baayen, R. H., Piepenbrock, R., & Van Rijn, H. (1993). The Celex Lexical Database
(CD-ROM). Linguistic Data Consortium, University of Pennsylvania,
Philadelphia, PA.
Brysbaert, M., Drieghe, D., & Vitu, F. (2005). Word skipping: Implications for
theories of eye movement control in reading. In G. Underwood (Ed.),
Cognitive Processes in Eye Guidance, Oxford: Oxford University Press, pp.
53-77.
Brysbaert, M., & Vitu, F. (1998). Word skipping: implications for theories of eye
movement control in reading. In G. Underwood (Ed.), Eye Guidance in
Reading and Scene Perception. Oxford: Elsevier, pp. 125-147.
Drieghe, D., Brysbaert, M., & Desmet, T. (2005). Parafoveal-on-foveal effects on eye
movements in text reading: Does an extra space make a difference? Vision
Research, 45, 1693-1706.
Drieghe, D., Brysbaert, M., Desmet, T., & Debaecke, C. (2004). Word skipping in
reading: On the interplay of linguistic and visual factors. European Journal of
Cognitive Psychology, 16, 79-103.
Drieghe, D., Rayner, K., & Pollatsek, A. (2005). Word skipping during reading
revisited. Journal of Experimental Psychology: Human Perception and
Performance, 31, 954-969.
Ehrlich, S. F., & Rayner, K. (1981). Contextual effects on word perception and eye
movements during reading. Journal of Verbal Learning and Verbal Behaviour,
20, 641-655.

21
Engbert, R., Longtin, A., & Kliegl, R. (2002). A dynamical model of saccade
generation in reading based on spatially distributed lexical processing. Vision
Research, 42, 621-636.
Engbert, R., Nuthmann, A., Richter, E. M., & Kliegl, R. (2005). SWIFT: A dynamical
model of saccade generation during reading. Psychological Review, 112, 777-
813.
Inhoff, A. W., & Rayner, K. (1986). Parafoveal word processing during eye fixations
in reading: effects of word frequency. Perception & Psychophysics, 40, 431 –
439.
Garnham, A., Oakhill, J., & Cruttenden, H. (1992). The role of implicit
causality and gender cue in the interpretation of pronouns. Language and
Cognitive Processes, 7, 231-255.
Garnham, A., Traxler, M., Oakhill, J., & Gernsbacher, M.A. (1996). The locus of
implicit causality effects in comprehension. Journal of Memory and Language,
35, 517-543.
Greene, S.B., & McKoon, G. (1995). Telling something we can't know: Experimental
approaches to verbs exhibiting implicit causality. Psychological Science, 6,
262-270.
Henderson, J. M., & Ferreira, F. (1993). Eye movement control during reading:
Fixation measures reflect foveal but not parafoveal processing difficulty.
Canadian Journal of Experimental Psychology, 47, 201-221.
Kerr, P. W. (1992). Eye movement control during reading: The selection of where to
send the eyes. Doctoral dissertation, University of Illinois at Urbana-
Champaign, Urbana, IL.
Kliegl, R., & Engbert, R. (2005). Fixation durations before word skipping in reading.
Psychonomic Bulletin & Review, 12, 132-138.
Liversedge, S. P., & Findlay, J. M. (2000). Saccadic eye movements and cognition.
Trends in Cognitive Sciences, 4, 6-14.

22
Long, D.L., & De Ley, L. (2000). Implicit causality and discourse focus: The
interaction of text and reader characteristics in pronoun resolution. Journal of
Memory and Language, 42, 545-570.
McDonald, J.L., & MacWhinney, B. (1995). The time course of anaphor resolution:
Effects of implicit verb causality and gender. Journal of Memory and
Language, 34, 543-566.
McKoon, G., Greene, S.B., & Ratcliff, R. (1993). Discourse models, pronoun
resolution, and the implicit causality of verbs. Journal of Experimental
Psychology: Learning, Memory, and Cognition, 19, 1040-1052.
Morrison, R. E. (1984). Manipulation of stimulus onset delay in reading: evidence for
parallel programming of saccades. Journal of Experimental Psychology:
Human Perception and Performance, 10, 667-682.
Radach, R., & Kempe, V. (1993). An individual analysis of initial fixation positions in
reading. In G. d’Ydewalle & J. Van Rensbergen (Eds.), Perception and
cognition: Advances in eye movement research (pp. 213-226). Amsterdam:
North Holland.
Radach, R., & Kennedy, A. (2004). Theoretical perspectives on eye movements in
reading: Past controversies, current issues, and an agenda for future research.
European Journal of Cognitive Psychology, 16, 3 – 26.
Rayner, K. (1978). Eye movements in reading and information processing.
Psychological Bulletin, 85, 618 – 660.
Rayner, K. (1979). Eye guidance in reading: Fixation locations within words.
Perception, 8, 21 – 30.
Rayner, K. (1998). Eye movements in reading and information processing: 20 years of
research. Psychological Bulletin, 124, 372 – 422.
Rayner, K., Binder, K. S., Ashby, J., & Pollatsek, A. (2001). Eye movement control in
reading: word predictability has little influence on initial landing positions in
words. Vision Research, 41, 943-954.

23
Rayner, K., & Duffy, S. A. (1986). Lexical complexity and fixation times in reading:
effects of word frequency, verb complexity, and lexical ambiguity. Memory &
Cognition, 14, 191 – 201.
Rayner, K., & Fischer, M. H. (1996). Mindless reading revisited: eye movements
during reading and scanning are different. Perception & Psychophysics, 58,
734-747.
Rayner, K., & McConkie, G. W. (1976). What guides a reader's eye movements?
Vision Research, 16, 829-837.
Rayner, K., Sereno, S. C., Morris, R. K., Schmauder, A. R., & Clifton, C. (1989). Eye
movements and on-line language comprehension processes. Language &
Cognitive Processes, 4, 1-50.
Rayner, K., Sereno, S. C., & Raney, G. E. (1996). Eye movement control in reading: a
comparison of two types of models. Journal of Experimental Psychology:
Human Perception and Performance, 22, 1188-1200.
Rayner, K., & Well, A. D. (1996). Effects of contextual constraint on eye movements
during reading: a further examination. Psychonomic Bulletin & Review, 3,
504-509.
Reichle, E. D., Rayner, K., & Pollatsek, A. (2003). The E-Z Reader model of eye
movement control in reading: comparisons to other models. Behavioral and
Brain Sciences, 26, 445 – 526.
Reilly, R. G., & Radach, R. (2003). Foundations of an interactive model of eye
movement control in reading. In Hyönä, J., Radach, R., & Deubel, H. (Ed.),
Cognitive and Applied Aspects of Eye Movement Research. Amsterdam:
Elsevier, pp. 429-455.
Reilly, R. G., & Radach, R. (2006). Some empirical tests of an interactive activation
model of eye movement control in reading. Cognitive Systems Research, 7,
34-55.
Schilling, H. E. H., Rayner, K., & Chumbley, J. I. (1998). Comparing naming, lexical

24
decision, and eye fixation times: word frequency effects and individual
differences. Memory & Cognition, 26, 1270 – 1281.
Starr, M. S., & Rayner, K. (2001). Eye movements during reading: some current
controversies. Trends in Cognitive Sciences, 5, 156-163.
Stewart, A. J., Pickering, M. J., & Sanford, A. J. (2000). The time course of the
influence of implicit causality information: Focusing versus integration
accounts. Journal of Memory and Language, 42, 423-443.
Vitu, F., O'Regan, J. K., Inhoff, A. W., & Topolski, R. (1995). Mindless reading: eye-
movement characteristics are similar in scanning letter strings and reading
texts. Perception & Psychophysics, 57, 352-364.
Vonk, W. (1984). Eye movements during comprehension of pronouns. In A. G. Gale
and F. Johnson (Eds.), Theoretical and Applied Aspects of Eye Movement
Research. Amsterdam: North-Holland. Pp. 203-211.
White, S. J. & Liversedge, S. (2004). Orthographic familiarity influences initial eye
fixation positions in reading. European Journal of Cognitive Psychology, 16,
52-78.

25
Table 1. Examples of the sentences used in the experiment as a function of pronoun gender
(masculine versus feminine) and gender cue (ambiguous versus unambiguous).
N1 Verb (“zich verontschuldigen” [apologize])
Masculine Pronoun - Ambiguous
Rik verontschuldigde zich bij Peter omdat hij een fout maakte.
[Rik apologized to Peter because he made an error.]
Masculine Pronoun - Unambiguous
Rik verontschuldigde zich bij Linda omdat hij een fout maakte.
[Rik apologized to Linda because he made an error.]
N1 verb (“verontrusten” [distress])
Feminine Pronoun - Ambiguous
Sarah verontrustte Ellen omdat zij vaak spijbelde.
[Sarah distressed Ellen because she skipped school often.]
Feminine Pronoun - Unambiguous
Sarah verontrustte Simon omdat zij vaak spijbelde.
[Sarah distressed Simon because she skipped school often.]

26
Table 2. Skipping probability of the pronoun and fixation times on the pronoun, the first-pass
time of the region after the pronoun and the total reading times for the entire sentences as a
function of gender cue and the gender of the pronoun.
Gender Cue No Gender Cue
hij (he) zij (she) hij (he) zij (she)
Skipping probability of the
pronoun 51 % 46 % 60 % 45 %
Single fixation time on the
pronoun (ms) 210 212 208 214
First pass time of the region
after the pronoun (ms) 352 410 340 394
Total reading time of the
sentences (ms) 2396 2489 2382 2497

27
Table 3. Mean fixation duration and launch site (distance from the pronoun in character
positions) prior to the skipping of or the landing on the pronoun as a function of gender cue and
the gender of the pronoun.
Gender Cue No Gender Cue
hij (he) zij (she) hij (he) zij (she)
Fixation duration prior to
skipping (ms) 223 225 220 226
Fixation duration prior to
landing (ms) 234 240 229 239
Launch site prior to
skipping 3.7 3.2 3.9 4.0
Launch site prior to landing
8.2 7.4 7.9 7.5