Orthographic familiarity influences initial eye fixation positions in reading

An important issue in the understanding of eye movements in reading is what kind of nonfoveal information can influence where we move our eyes. In Experiment 1, first fixation landing positions were nearer the beginning of misspelled words. Experiment 2 showed that the informativeness of word beginnings does not influence where words are first fixated. In both experiments, refixations were more likely to be to the left of the initial fixation position if the words were misspelled. Also, there was no influence of spelling on prior fixation durations or refixation probabilities, that is, there was no evidence for parafoveal‐on‐foveal effects. The results show that the orthographic familiarity, but not informativeness, of word initial letter sequences influences where words are first fixated.

Preprocessing of nonfixated text can also influence landing positions in reading. Studies have shown that landing positions are nearer the beginning of words or saccade lengths are shorter (1) when space but not letter (orthographic) information is presented (Inhoff, 1989;Morris et al., 1990;Rayner, Well, Pollatsek, & Bertera, 1982), (2) for words presented in uppercase letters (Inhoff, Starr, & Shindler, 2000b), (3) for words in which partially correct previews were presented before they were fixated (Inhoff, 1990), and (4) for nonsense letter strings presented in sentences (McConkie, Underwood, Zola, & Wolverton, 1985;Zola, 1984). The results of these studies indicate that nonfoveal processing beyond the level of visual word length information can influence where the eyes move. A number of studies have shown that lexical processing can influence where the eyes move in terms of the probability of fixating words (for a review of this area, see Brysbaert & Vitu, 1998). However, the issue to be investigated in this paper is what kinds of nonfoveal processing can influence where words are initially fixated. Some studies have suggested that nonfoveal processing of predictability within the context of the sentence (Lavigne, Vitu, & d'Ydewalle, 2000; but see Rayner, Binder, Ashby, & Pollatsek, 2001;Vonk, Radach, & van Rijn, 2000) and morphology (Inhoff, Briihl, & Schwartz, 1996;Hyo Èna È & Pollatsek, 1998; but see Andrews, Miller, & Rayner, 2004) can influence where words are first fixated. Although these studies suggest that nonfoveal lexical processing can influence fixation positions, word frequency has not been found to influence where words are first fixated (Rayner et al., 1996). Furthermore, the issue of whether nonfoveal sublexical processing can influence where words are first fixated remains controversial. The current paper focuses on how sublexical and lexical processing of nonfixated text can influence first fixation positions on words.
The first studies to investigate such landing position effects suggested that the distribution of informativeness (the importance of letter sequences for word recognition) within words could influence where words are initially fixated (Everatt & Underwood, 1992;Hyo Èna È, Niemi, & Underwood, 1989;Underwood, Clews, & Everatt, 1990;Underwood, Hyo Èna È, & Niemi, 1987). However, Hyo Èna È (1993Hyo Èna È ( , 1995 argued that these studies confounded the variable of informativeness with orthographic, morphological, and semantic variables. Furthermore, Rayner and Morris (1992) failed to replicate Underwood et al.'s (1990) findings. However, an increasing number of studies have shown that orthographic nonfixated information can influence where words are first fixated. Hyo Èna È (1995) presented words in Finnish sentences with either orthographically familiar or very unfamiliar word beginnings. The unfamiliar words included letters that rarely occur in Finnish. Hyo Èna È found that fixations landed nearer to the beginning of the orthographically unfamiliar words, especially on the space before the word. Radach, Inhoff, and Heller (2004) and Vonk et al. (2000) also found that orthographic familiarity influences where words are first fixated in German and Dutch respectively. In support of these findings, artificial task experiments 1 have shown that orthographically unfamiliar initial letter sequences produce first fixation landing positions nearer the beginning of French words and nonwords than letter strings with familiar initial letter sequences (Beauvillain & Dore Â, 1998;Beauvillain, Dore Â, & Baudouin, 1996;Dore Â & Beauvillain, 1997). However, a number of studies have failed to demonstrate orthographic landing position effects in artificial tasks (Kennedy, 1998(Kennedy, , 2000Radach, Krummenacher, Heller, & Hofmeister, 1995), short passage reading tasks (Liversedge & Underwood, 1998), and corpus reading studies (Radach & Kempe, 1993;Radach & McConkie, 1998).
A number of possible explanations have been proposed to explain the modulation of landing positions on words by orthography. Hyo Èna È (1993) suggested a pull assumption in which salient features (such as irregular orthography)``pop out'' of nonfixated text and pull the eye towards them. Beauvillain et al. (1996) extended this idea by suggesting that irregular letter sequences``pop out'' of nonfixated text and influence landing positions by adjusting the word length based saccade computation. More recently, Findlay and Walker (1999) suggested that medium-and long-term learning modifies the intrinsic salience of visual stimuli such as orthographic letter sequences. Intrinsic salience then contributes to a salience map in which the distribution of salience across the visual field determines the saccade target. Alternatively, the modulation of first fixation positions on words might be explained by an influence of processing difficulty. McConkie (1979) suggested that saccades are directed to regions of processing difficulty. Hyo Pollatsek (1998, 2000) suggested that processing difficulty reduces the perceptual span or the extent of preprocessing, and this shortens saccades (see also Hyo Èna È, 1995).
Overall, there is mixed evidence for whether sublexical and lexical processing can influence where words are first fixated. A similar controversy exists over the issue of whether nonfixated information can influence fixation durations, sometimes referred to as``parafoveal-on-foveal'' effects. A number of studies using both artificial and sentence reading tasks have reported that fixation durations are influenced by the characteristics of the following word (Inhoff, Radach, Starr, & Greenberg, 2000a;Inhoff et al., 2000b;Kennedy, 1998Kennedy, , 2000Kennedy, Pynte, & Ducrot, 2002;Murray, 1998;Murray & Rowan, 1998;Pynte, Kennedy, & Ducrot, 2004;Underwood, Binns, & Walker, 2000). However, the direction of the effects in these studies is inconsistent and there are concerns about the generalisability of some of the findings (Rayner, White, Kambe, Miller, & Liversedge, 2003).
The aim of this study was to investigate whether sublexical and lexical processing can influence where words are first fixated in the reading of English sentences. Experiment 1 was designed to test whether orthographic familiarity influences where the eyes move. Experiment 2 was designed to test whether processing of word beginnings to generate possible lexical candidates can influence where words are first fixated. Both experiments also provide an opportunity to test whether orthographic and lexical characteristics of nonfixated words influence fixation durations on the previous word. Throughout this paper we will use the term``orthographically unfamiliar'' to indicate that a letter sequence does not exist in the language or is infrequent. We will use the term`o rthographically familiar'' to indicate that a letter sequence frequently occurs in the language. Orthographic familiarity is measured using token frequency (the sum of the word frequencies of words including a particular letter sequence). We will use the term``informative'' to indicate that a letter sequence is found at the beginning of a small number of words and``uninformative'' to indicate that a letter sequence is found at the beginning of many words. Orthographic informativeness is measured using type frequency (the number of words including a particular letter sequence). In the present experiment, unless otherwise specified, type and token frequency counts were position specific and case insensitive.

EXPERIMENT 1
A correctly spelled condition with frequent word initial trigrams (e.g., agricultural) was compared to four misspelled conditions. Three of the misspelled conditions had different degrees of orthographic familiarity in order that we might identify those characteristics of an orthographically unfamiliar string that influence saccade computation. The most unfamiliar misspelling condition formed an illegal unpronounceable word initial trigram (e.g., ngricultural). The word``illegal'' indicates that no word in the English language begins with such an initial trigram. A second condition formed an illegal but pronounceable word initial trigram (e.g., akricultural). The third misspelled condition had word initial trigrams that occurred in the lexicon but which were unfamiliar (e.g., aoricultural). Following the evidence described above, it was predicted that these three misspelling conditions with unfamiliar initial letter sequences would produce first fixation landing positions nearer the word beginning compared to the correctly spelled condition.
The three unfamiliar misspelling conditions necessarily confounded the orthographic familiarity of the initial trigram with the presence of a misspelling. As a result, a fourth misspelling condition was included in which the word initial trigrams were frequent (e.g., acricultural). There were no differences in the initial bigram and trigram frequencies between the correctly spelled and high frequency misspelled conditions and so it was predicted that landing positions in these two conditions would be the same.

Method
Participants. Forty-five members of the University of Durham community participated in the experiment. All of the participants were native English speakers with normal or corrected to normal vision. The participants were paid to participate and all were naõ Ève in relation to the purpose of the experiment.
Materials. Word frequencies and n-gram frequencies were based on the CELEX English word form corpus (Baayen, Piepenbrock, & Gulikers, 1995). There were 35 critical strings with a mean word length of 9.7 characters (range 8±13) and a mean word frequency of 41 counts per million (SD = 53). Stimuli were chosen on the basis of position specific token frequency, but type frequencies followed similar patterns. N-gram frequencies were based on counts per 17.9 million because this is a more sensitive measure.
There were five spelling conditions that were manipulated within participants and items. In the baseline condition the critical string was spelled correctly with a high frequency initial trigram. In the high frequency, low frequency, and the illegal pronounceable misspelled conditions the second letter of the word was misspelled. In the illegal unpronounceable misspelled condition the first letter was misspelled if the original first letter was a vowel (21 items) and the second letter was misspelled if the original first letter was a consonant (14 items). There were 35 items in all of the conditions except for the illegal pronounceable misspelled condition in which there were 30 items. 2 The token frequencies for the initial trigrams were high for both the correctly spelled (M = 14,514, SD = 17,465) and the high frequency misspelled (M = 11,345, SD = 12,848) conditions. The stimuli were chosen primarily on the basis of token frequency but type frequencies for the initial trigrams also tended to be high for both the correctly spelled (M = 75, SD = 88) and the high frequency misspelled (M = 68, SD = 115) conditions. The position in the word at which the high frequency misspelled words became illegal was uncontrolled. The position ranged from four to seven characters and the mean position was 4.6 characters (SD = 0.88). In the low frequency misspelled condition the initial trigram mean type (M = 3, SD = 3) and token frequencies (M = 68, SD = 63) were low. To summarise, the position specific token frequencies of the initial trigrams were within the seventh to the ninth deciles for the correctly spelled condition, the sixth to the ninth deciles for the high frequency misspelling condition, and the first to the fifth deciles for the low frequency misspelling condition. For the illegal pronounceable and the illegal unpronounceable misspelled conditions the initial trigram never occurred in the lower case corpus. For the correctly spelled and the high frequency misspelled conditions there were no significant differences between the initial trigram type or token frequencies (ts < 1). There were significant differences in the initial trigram type and token frequencies between the correctly spelled and the high frequency misspelled conditions compared to the low frequency and illegal initial trigram conditions (ts > 3.3, ps < .01). The type and token initial bigram frequencies were high for the correctly spelled and high frequency misspelled conditions but low for the low frequency and illegal misspelled conditions. There were no significant differences in the initial or second (second and third letter) bigram type and token frequencies between the correctly spelled and the high frequency misspelled conditions (ts < 1.5). There were significant differences in initial bigram type and token frequencies between the correctly spelled and the high frequency misspelled conditions compared to the low frequency misspelled conditions and the illegal misspelled conditions (ts > 3.8, ps < .01). There were no significant differences in nonposition specific type or token monogram frequency between the correctly spelled and the high frequency misspelled condition (ts < 0.6) and the low frequency misspelled condition (ts < 1.9).
The critical strings in each condition were placed in identical sentence frames. Each of the sentences was no longer than one line of text (78 characters) and the critical strings appeared approximately in the middle of the sentence. The words before (word n±1) and after (word n+1) the critical string were either five or six letters long and had medium to high frequencies.
Five lists of 171 sentences were constructed and nine participants were randomly allocated to each list. Each list included 34 experimental sentences of which 6 items were from the illegal pronounceable condition and 7 items were from each of the other four conditions. The conditions were rotated following a Latin square design. There were 30 misspelled filler sentences with the misspellings in a variety of word lengths and positions. There were also 107 filler sentences that were spelled correctly. Therefore of the 171 sentences, 57 contained a misspelling. Fifty-eight of the sentences were followed by a comprehension question to ensure that participants concentrated on understanding the sentences. The sentences were presented in a fixed random order with eight filler sentences at the beginning. See the Appendix for examples of experimental sentences and critical words.
Procedure. Eye movements were monitored using a Dual Purkinje Generation 5.5 eye tracker. Viewing was binocular but only the movements of the right eye were monitored. The sentences were displayed on a screen at a viewing distance of 70 cm. Three and a half characters subtended one degree of the visual angle. The resolution of the eye tracker is 10 min of arc and the sampling rate was every millisecond.
Participants were instructed to ignore the misspellings and to concentrate on understanding the sentence to the best of their ability. A bite bar and head restraint were used to minimise head movements. The initial calibration procedure lasted approximately 5 min and the calibration accuracy was checked after every few trials during the experiment. After reading each sentence the participants pressed a button to continue and used a button box to respond`y es'' or``no'' to comprehension questions. The entire experiment lasted approximately 45 min and participants were given two breaks.
Analyses. Fixations shorter than 80 ms that were within one character of the next or previous fixation were incorporated into that fixation. Any remaining fixations shorter than 80 ms and longer than 1200 ms were discarded. Analyses of word n±1, the critical string and word n+1 included the space before the respective word in each case.
Five percent of trials were excluded due to either no first pass fixations on the sentence prior to word n±1 or tracker loss or blinks on first pass reading of word n±1 or the critical string. Five participants were replaced due to more than 15% of trials being excluded in this manner and one participant was replaced due to an error rate greater than 15% on the comprehension questions.

Results
Repeated measures analyses of variance (ANOVAs) were undertaken across the five spelling conditions, with participants (F 1 ) and items (F 2 ) as random variables. If there were significant main effects across both participants and items then simple effects were computed comparing the correctly spelled condition with each of the misspelled conditions. Paired samples t-tests were undertaken across both participants (t 1 ) and items (t 2 ) for comparisons between pairs of variables. The duration of the first fixation, gaze duration (the sum of fixations on a word before leaving it), and total time (the sum of all fixations within a word) were calculated for word n±1, the critical string and word n+1. For the critical string, landing positions, incoming saccade extent, launch site and frequency of refixations were analysed. The mean error rate on the comprehension questions was 5%.
Parafoveal-on-foveal effects. Table 1 shows the mean reading time measures on word n±1 and mean fixation durations prior to fixating the critical string. For word n±1 there were no significant effects of spelling on first fixation or gaze duration (Fs < 1.2). There was no effect of the spelling of the critical string on the probability of refixating word n±1 (Fs < 1). There was no significant effect of spelling on the duration of the fixation prior to the first fixation on the critical string for all of the data (Fs < 1), for only those trials when the prior fixation was on word n±1, F 1 (4, 176) = 1.42, p = .231, MSE = 1470; F 2 < 1, and for only those trials when the prior fixation was three or less characters away, 3 F 1 (4, 144) = 1.27, p = .284, MSE = 3556; F 2 < 1. These data provide no evidence of parafoveal-on-foveal effects.
There was no effect of spelling on the duration of the fixation after leaving the critical string (Fs < 1). There was no significant effect of spelling on the first fixation on word n+1 (Fs < 1) and although there was a significant effect of  TABLE 1 Reading time measures for each condition in Experiments (Exp) 1 and 2. Mean first fixation duration (FF) and gaze duration (GD) on word n±1, word n, and word n+1. Total time (TT) on word n and word n+1. Fixation duration prior to fixating the critical string (prior fixation) for all the data (All), for saccades launched from word n±1 (Word n±1) and saccades launched from three or less characters from the beginnings of the critical string (  spelling on gaze duration on word n+1 across participants this was not significant across items, 5 F 1 (4, 168) = 3.44, p = .01, MSE = 2309; F 2 (4, 116) = 1.93, p = .11, MSE = 2571. However, spelling did significantly influence total reading times on word n+1, F 1 (3.4, 151.1) = 5.31, p < .01, MSE = 5855; F 2 (4, 116) = 4.24, p < .01, MSE = 5646. The correctly spelled condition produced shorter total reading times on word n+1 than each of the misspelled conditions (Fs > 8.7, ps < .01).
To summarise, there were no effects of the critical string on prior fixations. However, fixation durations were increasingly longer on the critical string, and for later measures on word n+1, the more irregular the misspelling. The misspelled words, especially the very irregular misspellings, were more difficult to process than the correctly spelled words. Previous studies have also shown that misspelled words (Inhoff & Topolski, 1994;Rayner, Pollatsek, & Binder, 1998b;Underwood, Bloomfield, & Clews, 1988), and words that are incorrect in the context of the sentence (Daneman, Reingold, & Davidson, 1995;Ehrlich & Rayner, 1981) produce longer fixation durations.
Landing positions. For the first fixation landing position analyses, the space before the critical string was classified as zero and the first letter of the string as one, etc. Table 2 shows the mean landing positions on the critical string. The mean first fixation landing positions on the critical string were 0.5 to 0.3 characters nearer the word beginning for the misspelled strings compared to the correctly spelled words. There was a significant effect of spelling on the mean first fixation landing position on the critical string, F 1 (4, 176) = 3.08, p = .02, MSE = 0.75; F 2 (4, 116) = 2.91, p = .02, MSE = 0.57. Compared to the correctly spelled condition, mean landing positions were significantly nearer the beginning of the word in the high frequency, F 1 (1, 44) = 7.46, p < .01, MSE = 1.6; F 2 (1, 29) = 15.14, p < .01, MSE = 0.69, low frequency, F 1 (1, 44) = 8.21, p < .01, MSE = 1.01; F 2 (1, 29) = 4.97, p = .03, MSE = 1.08, illegal pronounceable, F 1 (1, 44) = 14, p < .01, MSE = 1.07; F 2 (1, 29) = 5.7, p = .02, MSE = 1.54, and the illegal unpronounceable, F 1 (1, 44) = 4.07, p = .05, MSE = 1.3; F 2 (1, 29) = 6.1, p = .02, MSE = 1.17 misspelled conditions. No other paired contrasts between the misspelled conditions were significant (ts < 1.4). Figure 1 shows the distribution of landing positions for each condition. For all of the conditions, most fixations landed on the preferred viewing position (between the middle and the beginning of the word). The correctly spelled condition landing position distribution curve is shifted to the right of the misspelled condition curves. Clearly, readers processed the critical string prior to direct fixation and misspellings produced 5 In Experiment 1, the F 1 analysis of first pass reading time on word n+1 was based on data from 43 participants due to skipping and excluded data.
landing positions nearer to the beginning of the critical string compared to when it was spelled correctly.
Incoming saccade extent and launch site. Table 2 shows the mean lengths of saccades into the critical string and mean launch sites. There was no significant effect of spelling on the length of the saccade into the critical string, F 1 (4, 176) = 1.35, p = .253, MSE = 1.06; F 2 < 1. However, Table 2 shows that the mean saccade lengths into the critical string were longer for the correctly spelled condition than in any of the misspelled conditions. There was no effect of spelling on the position of the fixation prior to first fixating the critical string (Fs < 1) and, in contrast to the means for saccade lengths, Table 2 shows no consistent pattern of differences in launch site between the correctly spelled condition and the misspelled conditions. The probability of skipping word n±1 before fixating the critical string (trials in which regressions were made from word n±1 were considered separately) was numerically greater for the misspelled conditions (high frequency: 0.17; low frequency: 0.11; illegal pronounceable: 0.14; illegal unpronounceable: 0.16) than for the correctly spelled condition (0.09). There was a significant main effect of spelling on the probability of skipping word n±1, F 1 (3.2, 141.8) = 3.55, p = .01, MSE = 202; F 2 (4, 116) = 3.26, p = .01, MSE = 154, and word n±1 was less likely to be skipped in the correctly spelled condition than the high frequency, F 1 (1, 44) = 10.5, p < .01, MSE = 328; F 2 (1, 29) = 11.93, p < .01, MSE = 225, and illegal unpronounceable, F 1 (1, 44) = 9.07, p < .01, MSE = 261; F 2 (1, 29) = 4.83, p = .04, MSE = 447, misspelled conditions. There was no significant difference for the low frequency misspelling condition, Fs < 1.2, and a significant difference across participants, F 1 (1, 44) = 5.69, p = .02, MSE = 172, but not items, F 2 (1, 29) = 2.33, p = .138, MSE = 312, for the illegal pronounceable misspelling condition. The finding that spelling influenced the probability of skipping the previous word suggests that the misspellings may have attracted saccades from distant launch sites. Figure 2 shows the distribution of launch sites for each condition. Note that there tend to be slightly more saccades launched from seven or more characters from the critical string for the high frequency and illegal unpronounceable misspelled conditions compared to the correctly spelled condition. Therefore, although there were no significant effects of incoming saccade extent or launch site, the results suggest that differences in both these factors may have contributed to the landing position effect. That is, nonsignificant differences in the mean incoming saccade extent and differences in the probability of skipping word n±1 may both have influenced first fixation positions on the critical string.

FIXATION POSITIONS ON MISSPELLED WORDS
Frequency of refixations. Table 3 shows the probability of skipping or making one or more than one fixation on the critical string on first pass. There was a significant effect of spelling on the probability of refixating the critical string on first pass, F 1 (4, 176) = 8.21, p < .01, MSE = 307; F 2 (4, 116) = 6.73, p < .01, MSE = 287. The correctly spelled condition was significantly less likely to be refixated than any of the four misspelled conditions (Fs > 4.6, ps < .05). Similar to the reading time measures, these results suggest that the misspelled words were more difficult to process than the correctly spelled words because they produced more first pass refixations on the critical string. Table 3 shows the probability of making a refixation to the left of the initial fixation on the critical string for those trials in which multiple first pass fixations occurred on the critical word. There was a significant effect of spelling on the probability of making a refixation to the left of the initial fixation, 6 F 1 (4, 104) = 2.77, p = .03, MSE = 1003; F 2 (4, 88) = 5.91, p < .01, MSE = 666. Compared to the correctly spelled condition, refixations were more likely to be to the left of Figure 2. Launch site distributions for each condition for Experiment 1. 6 In Experiment 1, for the analyses of the probability of refixating to the left, only participants and items that produced refixations in all of the conditions were included. Consequently the F 1 analysis was based on 27 participants and the F 2 analysis was based on 23 items. the initial fixation in the low frequency, F 1 (1, 26) = 5.81, p = .02, MSE = 2308; F 2 (1, 22) = 7.6, p = .01, MSE = 1434, and the illegal unpronounceable, F 1 (1, 26) = 8.03, p < .01, MSE = 2134; F 2 (1, 22) = 18.02, p < .01, MSE = 1255, misspelled conditions. For the illegal pronounceable misspelled condition the effect was significant across participants but not items, F 1 (1, 26) = 4.27, p = .05, MSE = 1412; F 2 (1, 22) = 2.82, p = .107, MSE = 1490. There was no significant difference in the probability of refixating to the left of the initial fixation for the correct and high frequency misspelled words, F 1 (1, 26) = 1.27, p = .27, MSE = 1848; F 2 < 1. The results suggest that the more irregular misspelled items were more likely to produce regressive refixations than the correctly spelled condition.

Discussion
Both the position of the first fixation and the direction of refixations were influenced by the presence of misspellings. Therefore the results show that nonfoveal processing at least at the level of orthographic familiarity can influence where words are fixated. It is also interesting that the word before the critical word was more likely to be skipped if the critical word was misspelled. These results suggest that misspellings can attract saccades from distant launch sites. However, note that in contrast to these results, Pynte et al. (2004) found that irregular misspellings reduce the probability of skipping the previous word.
The effects of the misspellings on landing positions clearly show that they must have been processed during the previous fixation. However, importantly, there was no effect of the misspellings on prior fixation durations or refixation probabilities. That is, there were no parafoveal-on-foveal effects.

FIXATION POSITIONS ON MISSPELLED WORDS
First fixation landing positions were also significantly nearer the word beginning for the high frequency misspelled condition compared to the correctly spelled condition. There were no differences between initial trigram frequencies in the correct and high frequency misspelling conditions and, as mentioned above, the position in the word at which the high frequency misspelled words became illegal was uncontrolled. Therefore letter sequences up to at least the first four letters of the high frequency misspelled strings had to be processed in order for the word initial letter sequence to be detected as irregular. Importantly, although the string would have to be processed up to at least the fourth letter, shorter infrequent or illegal letter sequences may have been detected (for example quc in equcation). Nevertheless, this result is surprising because Beauvillain and Dore Â (1998) showed that the letter sequence frequency of the second and third letters of a letter string did not influence fixation positions. The high frequency misspelling result suggests that letter sequences that are positioned further into the word than those tested by Beauvillain and Dore Â can influence fixation positions. The high frequency misspelling result is even more surprising because it appears to be at least partly determined by saccades being attracted from distant launch sites. That is, word n±1 is more likely to be skipped when there is a high frequency misspelling than when the critical string is spelled correctly. Note that saccades launched from distant launch sites tend to land nearer the beginning of words (McConkie et al., 1988), hence the numerically larger number of fixations on the space before the word for the high frequency misspellings (see Figure 1). These results imply that processing beyond the level of the orthographic familiarity of the word initial trigram can be undertaken on text presented more than one word from fixation, which is beyond or certainly towards the far edge of the region of text from which letter information can be processed (McConkie & Rayner, 1975;Rayner, 1975. There are three possible explanations for the high frequency misspelling landing position result. First, the high frequency misspelled strings, or letter sequences within these strings, might have been identified as illegal. Second, the stimuli were chosen primarily on the basis of token, rather than type, frequency. Although there were no significant differences in type frequency, the high frequency misspelling condition had initial trigrams with a greater range and variation (range: 1±679, SD = 115) than the correctly spelled condition (range: 8±421, SD = 88) and therefore differences in type frequency (i.e., informativeness) might have influenced fixation positions. Third, given that the high frequency misspelling result was a surprising finding in relation to previous research, it may be spurious.

EXPERIMENT 2
Experiment 2 was designed to distinguish between the three alternative explanations for the high frequency misspelling result in Experiment 1. In order to do this, we constructed a set of materials in which a critical string was spelled correctly with an uninformative initial trigram, misspelled with an uninformative initial trigram, or misspelled with an informative initial trigram. Each of the three explanations for the results of Experiment 1 described above generate different predictions for Experiment 2. First, if fixation positions are influenced by any kind of misspelling then landing positions should be nearer the beginning of the critical string in both of the misspelled conditions, compared to the correctly spelled condition. Second, if lexical preprocessing of a nonfixated word influences where words are first fixated then the informativeness of word beginnings might affect landing positions. Consequently, we might expect fixation positions to be nearer the beginning of the informative misspelled strings compared to the uninformative misspelled and correctly spelled strings. Third, if processing beyond the level of the orthographic familiarity of word initial trigrams does not influence where words are first fixated then there should be no difference in landing positions for letter strings with equally familiar initial letter sequences. That is to say, there should be no effect of spelling on landing positions.
Previous studies of the effects of orthographic regularity on landing positions (Beauvillain & Dore Â, 1998;Beauvillain et al., 1996;Dore Â & Beauvillain, 1997;Everatt & Underwood, 1992;Hyo Èna È, 1995;Hyo Èna È et al., 1989;Radach et al., 2004;Underwood et al., 1987Underwood et al., , 1990Vonk et al., 2000) have confounded the variables of orthographic familiarity and informativeness. Kennedy (2000) did carefully manipulate these, and other, variables in artificial task experiments but found no effects of letter sequence frequency on landing positions. Experiment 2 provides a test of the hypothesis that informativeness of the word initial trigam, independent of the orthographic familiarity of the word initial trigram, influences fixation positions.

Method
Participants. Twenty-four native English speakers at the University of Durham were paid to participate in the experiment. The participants all had normal or corrected to normal vision and were naõ Ève in relation to the purpose of the experiment.
Materials. Word frequencies and n-gram frequencies were calculated using the CELEX English word form corpus (Baayen et al., 1995). All of the critical words were eight or nine characters long (M = 8.5, SD = 0.5) and the mean word frequency in counts per million was 17 (SD = 29.5). There were three conditions that were manipulated within participants and items. The critical words were spelled correctly or the second letter was misspelled to create either an uninformative or informative initial trigram.
Position specific n-gram frequencies were calculated in counts per 17.9 million. The initial trigram token frequencies tended to be higher in the uninformative (M = 8696, SD = 10,545) and informative (M = 6879, SD = 10,974) misspelled conditions compared to the correctly spelled condition (M = 4325, SD = 5262), although these differences were not significant (ts < ±1.8, ps > .9). The uninformative misspelled condition also tended to have higher type frequency initial trigrams (M = 47, SD = 26) compared to the correctly spelled condition (M = 32, SD = 29) although this was not significant, t(23) = ±1.72, p = .1. Importantly, the informative misspelled condition had significantly lower type frequency initial trigrams (M = 6, SD = 3) than both the correctly spelled condition, t(23) = 4.45, p < .01 and the uninformative misspelled condition, t(23) = 7.51, p < .01. The type and token frequencies of the initial bigrams followed a similar pattern. The type and token monogram frequencies of the misspelled second letter were significantly higher in the misspelled condition compared to the correctly spelled condition (ts > 2). Importantly, the initial trigrams of the misspelled conditions were not more orthographically unfamiliar than the correctly spelled condition.
The 24 critical words were embedded in identical sentence frames for each condition. Each of the sentences was no longer than one line of text (78 characters) and the critical word appeared approximately in the middle of the sentence. The words before and after the critical word were either five or six letters long and had medium to high frequencies. Most of the sentences included context relevant to the critical word at the beginning of the sentence. See the Appendix for example experimental sentences and critical words.
Three lists of 96 items were constructed and eight participants were randomly allocated to each list. Each list included 24 experimental items of which 8 items were from each of the three misspelling conditions. The conditions were rotated following a Latin square design. There were 16 misspelled filler items with misspellings in a variety of word lengths and in a variety of positions within the word and the sentence. There were also 56 filler items that were spelled correctly. Therefore, of the 96 items, 32 contained a misspelling. Thirty-two of the sentences were followed by a comprehension question to ensure that participants concentrated on understanding the sentences. The sentences were presented in a fixed random order with six filler sentences at the beginning.
Procedure. The experimental procedure was the same as in Experiment 1. The entire experiment lasted approximately 30 min and participants were given one break.
Analyses. The analyses were the same as in Experiment 1. Of the trials, 1.4% were excluded due to either no first pass fixations on the sentence prior to word n±1 or tracker loss or blinks on first pass reading of word n±1 or the critical string.

Results
The results were analysed in the same manner as for Experiment 1. The mean error rate on the comprehension questions was 2.3%.
Parafoveal-on-foveal effects. Table 1 shows the mean fixation durations prior to fixating the critical string. There were no significant effects of spelling on first fixation, gaze duration, or total time for word n±1 (Fs < 1.1). There was also no difference in the probability of refixating word n±1 (Fs < 1). There were no significant effects of spelling on the duration of the fixation prior to first fixating the critical word for all of the data, for saccades launched from word n±1 and for saccades launched from three or less characters from the critical word (Fs < 1). Once again, the results show no evidence of parafoveal-on-foveal effects.
Reading times were longer on the misspelled words than the correctly spelled words and there were no reliable spillover effects. These data indicate that misspelled words were more difficult to process than the correctly spelled words.
Landing position. Table 2 shows the mean first fixation positions on the critical string in each of the conditions. There were no significant effects of spelling on landing position (Fs < 1). Figure 3 shows the distribution of landing positions for each of the conditions; note that most fixations landed on the preferred viewing position.
Incoming saccade extent and launch site. Table 2 shows the mean saccade extents and launch sites for each condition. There were no effects of spelling on the launch site or incoming saccade extent prior to fixating the critical string (Fs < 1). There was no effect of spelling on the probability of skipping word n±1 before fixating the critical string (trials in which regressions were made from word n±1 were considered separately) (Fs < 1). Table 3 shows the probability of skipping or making one or more than one fixation on the critical string on first pass. There was a significant effect of spelling on the probability of refixating the critical string on first pass, F 1 (2, 46) = 9.7, p < .01, MSE = 211; F 2 (2, 46) = 6.69, p < .01, MSE = 316. The correctly spelled condition was significantly less likely to be refixated on first pass than the uninformative misspelling condition, F 1 (1, 23) = 9.88, p < .01, MSE = 508; F 2 (1, 23) = 11.51, p < .01, MSE = 471, or the informative misspelling condition, F 1 (1, 23) = 16, p < .01, MSE = 443; F 2 (1, 23) = 10.65, p < .01, MSE = 672. Similar to the reading time measures, these results suggest that the misspelled words were more difficult to process because they produced more first pass refixations.

Discussion
Experiment 2 clearly showed that neither the presence of a misspelling nor the informativeness of word initial letters reliably influenced where words were first fixated. The results do not support the suggestion that lexical factors influence first fixation positions on words. There is a clear inconsistency between the high frequency misspelling effect in Experiment 1 and the results of Experiment 2. It must be concluded that influences on landing positions beyond the level of the orthographic familiarity of the initial trigrams are, at best, unreliable. However, the results do show that misspellings influence the directions of refixations.
Similar to Experiment 1, the results also provide no support for the notion that nonfixated text influences fixation durations. There was also no difference in reading times on words that were misspelled with an informative or uninformative initial trigram. These results suggest that the time to process a misspelled word is not influenced by the number of possible candidates that are generated by the initial trigram.

GENERAL DISCUSSION
The aim of this study was to investigate whether orthographic familiarity and informativeness influence landing positions in the reading of English sentences. In accordance with previous evidence, our results clearly show that most fixations land left of the word centre on the preferred viewing position (Dunn-Rankin, 1978;McConkie et al., 1988;Rayner, 1979) and that more distant launch sites are associated with landing positions nearer the beginning of words (Hyo Èna È, 1995;McConkie, Kerr, & Dyre, 1994;McConkie et al., 1988;Radach & Kempe, 1993;Radach & McConkie, 1998;Rayner et al., 1996). The results of the present study provide no reliable support for the influence of lexical processing on fixation positions. The most important finding is that first fixation landing positions were nearer the beginning of words with unfamiliar orthography. To our knowledge, this is the first study to find an effect of orthographic familiarity on landing positions for English language sentences. The results are supported by other recent experiments that have shown that orthographic processing influences where words are first fixated in the reading of correctly spelled English sentences . These results are particularly striking since current models of eye movements in reading make no attempt to account for the influence of orthography on first fixation landing positions (O'Regan, 1990;Reichle et al., 1998Reichle et al., , 1999Reilly & O'Regan, 1998). Our experiments are a clear demonstration that these effects do occur and any complete model of oculomotor control in reading must account for them.
The results of Experiment 1 support a number of studies in languages other than English that have shown that word initial orthographic regularity influences where words are first fixated (Beauvillain & Dore Â, 1998;Beauvillain et al., 1996;Dore Â & Beauvillain, 1997;Hyo Èna È, 1995;Radach et al., 2004;Vonk et al., 2000). However, in support of Rayner and Morris (1992), and against the claims of Underwood and colleagues (Everatt & Underwood, 1992;Hyo Èna È et al., 1989;Underwood et al., 1987Underwood et al., , 1990, the results provide no reliable support for the suggestion that informativeness, or the number of potential lexical candidates, can influence where words are first fixated. Experiment 1 provided a strong test of the possibility that orthographic familiarity produces parafoveal-on-foveal effects and Experiment 2 provided a strong test of the possibility that informativeness produces parafoveal-on-foveal effects. However, both experiments showed no evidence of such effects. Previous studies using artificial tasks (Kennedy, 1998(Kennedy, , 2000Murray, 1998;Murray & Rowan, 1998), and sentence reading studies (Inhoff et al., 2000a(Inhoff et al., , 2000bKennedy et al., 2002;Pynte et al., 2004;Underwood et al., 2000) have suggested that the characteristics of a word can influence fixation times on the previous word. However, as also reported in Rayner et al. (in press), in both the experiments reported here, there were no effects of the spelling of the critical string on the fixation duration prior to fixating the critical string, even when the prior fixation was three or less characters from the beginning of the critical string. Furthermore, the landing position effects in Experiment 1 show that the initial letter sequences were processed before the critical strings were fixated and yet there was no effect of the initial letter sequence on the prior fixation duration. Importantly, the absence of such effects in the present study at least suggests that there is no strong consistent parafoveal-on-foveal processing in normal sentence reading. That is, the results provide no support for the hypothesis that words are processed in parallel such that the sublexical and lexical characteristics of a nonfixated word can influence fixation durations or refixation probabilities on the previous word. Clearly, these data indicate that nonfoveal processing influences where the eyes move but does not influence when the eyes move to fixate a nonfoveal string.
Three types of possible explanations have been proposed to account for the effects of orthographic familiarity on fixation positions found in Experiment 1. First, Hyo Èna È (1993) and Beauvillain and Dore Â (1998) suggested that irregular letter sequences``pop out'' and attract saccades towards them. Second, Findlay and Walker (1999) proposed that visually unfamiliar (intrinsically salient) letter sequences could influence saccade computation within a salience map. The third possible explanation is that nonfoveal processing difficulty reduces the perceptual span and consequently shortens saccades (Hyo Èna È, 1995;Hyo Èna È & Pollatsek, 1998. In order for the processing difficulty hypothesis to explain the results of Experiment 2, it must be assumed that the uninformative and informative misspellings did not induce sufficient nonfoveal processing difficulty to influence saccades. A further important question is whether the effects of orthography on landing positions form part of a strategy that is efficient for word recognition, or whether they are simply due to mandatory processes that function in all types of visual processing such as scene viewing or visual search. Legge, Klitz, and Tjan (1997) suggested that an ideal reader would use all available information to determine the length of a saccade in order to minimise uncertainty for word recognition. However, if this was the case then it is not clear why a number of studies have failed to find effects of within word characteristics on first fixation positions. The landing position differences in Experiment 1 were very small and one might reasonably assume that such differences would be much greater if landing positions were central to visual word recognition. Also, landing position effects such as increasing the probability of landing on the space before the word are not necessarily optimal for word recognition (McConkie, Kerr, Reddix, Zola, & Jacobs, 1989;Vitu, O'Regan, & Mittau, 1990). If the influence of orthography on landing positions is not a strategy for efficient word recognition then an alternative possibility is that the effects might simply result from standard visual processes, as suggested by Findlay and Walker (1999).
The current experiments also yielded interesting results regarding the nature of refixations. O' Regan (1990) argued that the locations of refixations are determined by the position of the first fixation on a word in relation to the word length. However, in both Experiments 1 and 2 there were more refixations to the left for misspelled critical strings. In support of previous studies (Hyo Èna È, 1995;Hyo Èna È & Pollatsek, 1998;Pynte, 1996Pynte, , 2000Pynte, Kennedy, & Murray, 1991) the results suggest that the characteristics of a word influence the location of refixations. Once again, current models of eye movements in reading make no attempt to explain these effects.
In conclusion the results show that the orthographic familiarity, but not informativeness, of nonfixated words influences where words are first fixated and refixated. However, importantly the results provide no evidence for parafoveal-on-foveal effects.
PrEview proof published online September 2003