Is RoAsT Tougher than StEAk ? The Effect of Case Mixing on Perception of Multi-letter Graphemes

Case mixing is a technique that is used to investigate the perceptual processes involved in visual word recognition. Two experiments examined the effect of case mixing on lexical decision latencies. The aim of these experiments was to establish whether different case mixing patterns would interact with the process of appropriate visual segmentation and phonological assembly in word reading. In the first experiment, case mixing had a greater effect on response times to words when it led to visual disruption of the multi-letter graphemes (MLGs) as well as the overall word shape (e.g. pLeAd), compared to when it disrupted overall word shape only (e.g. plEAd). A second experiment replicated this finding with words in which MLGs represent either the vowel (e.g. bOaST vs. bOAst) or the consonant sound (e.g. sNaCK vs. sNAcK). These results confirm that case mixing can have different effect depending on the type of orthographic unit that is broken up by the manipulation. They demonstrate that graphemes are units that play an important role in visual word recognition, and that manipulation of their presentation by case mixing will have a significant effect on response latencies to words in a lexical decision task. As such these findings need to be taken into account by the models of visual word recognition.

involves presenting letter strings in which the case of the letters alternates within the string (e.g.wOrD).It has repeatedly been shown that word recognition is slower for mixed-case stimuli than for those in which all letters within a string are presented in the same case (e.g.Coltheart & Freeman, 1974;Allen, Wallace & Weber, 1995;Besner & McCann, 1987;Herdman, Chernecki & Norris, 1999;Mayall & Humphreys, 1996;Mayall, Humphreys & Olson, 1997).
Most accounts of the disruption caused by case mixing agree that the interference occurs at an early stage in word recognition, when visual features are encoded.However, there is less agreement about the nature of the features that might be affected.Mayall et al. (1997) identified four possible explanations of the case mixing effect, based on: disruption of word shape, disruption of trans-letter features (e.g.distortion of the shape of the spaces between letters), lateral masking of small by capital letters, and inappropriate grouping of same case letters.The results of experiments in which letter spacing and size were manipulated independently of case led them to conclude that both trans-letter features and the visual grouping of letters are important in word recognition.
The analysis of spelling patterns in terms of graphemes is necessary because written English does not preserve a one-to-one correspondence between letters and phonemes (e.g.Coltheart, 1978;Venezky, 1970).The term grapheme refers to the written representation of a phoneme, which can consist either of a single letter or a group of letters.This definition means that we can distinguish between single-letter graphemes -e.g. the letter u in grunt; and multi-letter graphemes (MLGs) -e.g. the letter pair oa in roast.The existence of MLGs may imply that phonological assembly requires an initial parsing of the letter string, in which some letters are grouped as multi-letter graphemes.
In a nonword naming study by Rastle & Coltheart (1998), naming latencies for five letter nonwords with three phonemes were longer than naming latencies for five letter nonwords with five phonemes.The interpretation offered by Rastle & Coltheart was that, during phonological assembly, the GPC system translates letters from a string onto phonemes serially, from left to right.It is important to note that the crucial unit in this account of conversion from orthography to phonology is a letter rather than grapheme.Therefore, when the GPC system encounters a multi-letter grapheme it initially starts by translating just the first letter of the grapheme, giving rise to incorrect phonological activation which inhibits activation of the correct phonological form.This competitive activation of spurious phonemes during processing of MLGs was described by Rastle & Coltheart as a "whammy effect" and it slows down activation of the correct phoneme leading to the latency cost.
The interpretation of the phonemic length effect offered by Rey et al. (1998) is somewhat different.Rey et al. (1998) compared five letter monosyllabic words containing three or five phonemes in a perceptual identification task, and obtained longer identification times for words with fewer phonemes.This effect was observed for both English and French low frequency words, but not for French high frequency words.According to Rey et al. (1998), these results indicate that grouping letters into graphemes requires additional processing time.The hypothesis that MLGs are processed as perceptual units in reading was further supported by the results of Rey et al. (2000) letter detection study which demonstrated that, in English and French, it takes longer to detect a letter when it is embedded in a MLG.Pring (1981) addressed the issue of the "psychological reality" of graphemes by investigating the effect of case mixing on the pseudohomophone (PH) effect.The PH effect refers to the finding that, in a lexical decision task, it takes longer to reject nonwords which can be pronounced as a word (e.g.brane), than other nonwords.Pring found that the PH effect is abolished under certain patterns of case mixing.If case mixing disrupts MLGs (e.g.CheRCH) H H the PH effect is eliminated, but, if MLGs are left intact by the case mixing (e.g.CHerCH) the PH effect is still obtained.She takes these results as evidence for H H the psychological reality of MLGs and their involvement in the construction of a nonlexical phonological code.
While it has generally been accepted that phonological assembly plays a role in nonword reading, its role in word recognition is often regarded as less important.This raises the question of whether MLGs have an important role in word recognition as well as in nonword recognition, and whether different case mixing patterns will interact with the process of appropriate visual segmentation and phonological assembly in word reading.If processing of MLGs requires additional processing time under normal (i.e.same case) presentation, then disruption of their processing by case mixing should cause even greater processing delays as inappropriately segmented MLGs impair correct perceptual grouping of graphemes and activate competing phonological elements.

EXPERIMENT 1
The aim of Experiment 1 was to investigate the effect of case mixing on the recognition of words containing MLGs; specifically, to determine how reading is affected when case changes coincide or conflict with the boundaries of multi-letter grouping.The hypothesis outlined above leads to the prediction that case mixing will have a greater impact on reading words in which MLGs are disrupted by case changes (e.g.rOaSt), than for those in which MLGs are preserved (e.g.rOAst).
Any direct comparison of these experimental conditions is open to the criticism of confounding type of case mixing with the number of case alternations within a letter string, as there are fewer case alternations in rOAst than in rOaSt.If there is a reading advantage for the case mixing pattern with intact MLGs, the effect could be due to the smaller number of case alternations, which reduces the degree of visual disruption.In an attempt to control for the possible effect of number of case alterations, a set of words consisting entirely of single-letter graphemes was also included in the experiment.For these words, the patterns of case mixing were the same as for words containing MLGs (e.g.tRUmp vs. tRuMp).

Method
Participants: Twenty students participated in the experiment as part of a course requirement.All participants were native speakers of English, and had normal or corrected-to-normal vision.
Stimuli and design: The experiment followed a three factor, within group design.The three independent variables manipulated in this experiment were: lexicality of items (word vs. nonword), type of vowel grapheme (single-vs.multi-letter grapheme) and pattern of case mixing ("disruptive pattern" which disrupts MLGs vs. "non-disruptive pattern" which leaves MLGs intact).
A total of 40 five letter monosyllabic words were selected for the experiment.Of these, 20 contained single-letter vowel graphemes (e.g.grunt), and 20 contained multi-letter vowel graphemes (e.g.roast).These two sets of words were matched on frequency.An additional 40 nonwords used in the experiment were derived from the selected words by changing the identity of the vowel.All items, together with their response times and error rates, are presented in Appendix 1.An additional set of monosyllabic words and nonwords were used for 24 practice trials.
Out of the 20 words with MLGs, 10 were presented with a case mixing pattern that had 4 case changes and disrupted MLGs (e.g.rOaSt), while for the other 10, the case mixing pattern had 2 case changes, leaving MLGs intact (e.g.hOUnd).The two patterns of case d d mixing were matched in the 20 words with single-letter vowel graphemes (e.g.gRuNt and cLAmp).These patterns of case mixing for each word were counterbalanced during the experiment.Forty nonwords used in the experiment were presented with the same patterns of case mixing as their corresponding words.The sequence of trials in the experimental session was randomised.Practice trials had the same pattern of case mixing as experimental trials.
Procedure: The experiment was based on the lexical decision task.Participants were tested individually in a quiet room and were seated approximately 60 cm from the computer monitor.Participants were told that they would be seeing a series of words or nonwords on the monitor, one at a time.They were instructed to respond by pressing a "yes" button on a response box if the stimulus was a real word, or a "no" button if it was a nonword, and to try to do this as quickly and accurately as possible.Each trial began with a 500 ms presentation of a fixation point in the centre of the screen; this was followed immediately by the presentation of the target stimulus, which stayed on screen for 1500 ms, or until the response was made.When one of the buttons was pressed, the stimulus would disappear from the screen.Response time was measured as the time between the onset of a stimulus and the pressing of the button.
The experimental session was preceded by 24 practice trials with the same type of task in order to familiarise participants with the procedure.The experimental session was divided into two blocks of 40 trials with a short break between them.

Results
Response times for each participant were trimmed to exclude trials where the response latency was more than two standard deviations from the mean for that participant.Mean response times and error rates for the remaining word and nonword trials in each condition in the by-participant analysis are shown in Table 1.For the analysis by stimuli, the trimming procedure excluded response times which were more than two standard deviations above or below the mean for a given stimulus.In addition to trimming data points that were two standard deviations above or below the mean, items associated with a high proportion of incorrect responses (30% or more) were excluded from the analysis.There were three such items; one was a word with MLG vowels, and the other two contained single letter vowel graphemes.
Analysis of word trials.In the analysis of word trials, the experiment was analysed as a two factor design, with type of vowel grapheme and type of case mixing as independent variables In an analysis of variance by participants (F1), type of vowel grapheme and type of case mixing were treated as within group factors.In the corresponding analysis of variance by stimuli (F2), type of vowel grapheme was treated as a between group factor, and type of case mixing as a within group factor.For response times, the main effect of the type of vowel grapheme was significant in the analysis by participants, F1(1,19)=4.96,MSE=2398, p<0.05), but not in the analysis by stimuli, with words in which the vowel was represented by MLG recognised faster than words containing single letter vowel graphemes.The main effect due to type of case mixing was not significant in either of the analysis.However, there was a significant two-way interaction between type of vowel grapheme and type of case mixing, F1(1,19)=4.79,MSE=4569, p<0.05;F2(1,35)=7.3,MSE=4074, p<.01.A planned comparison revealed a significant difference between the two case mixing patterns for words with MLG vowels, F1(1,19)=7.46,MSE=3544, p<0.01;F2(1,35)=9.92,MSE=4074, p<0.01.Words in which MLGs were preserved by the case mixing pattern were recognised faster than words in which the MLGs were disrupted.On the other hand, the pattern of case mixing had no significant effect on response times for words with single letter vowel graphemes.
When these analyses were repeated for error rates there was a significant main effect of the type of vowel in the analysis of variance by participants, F1(1,19)=7.18,MSE=0.004,p<0.01, but not in the analysis by stimuli.The main effect of type of case mixing and the two-way interaction between type of vowel grapheme and type of case mixing were not significant in either analysis.
Analysis of nonword trials.In the analysis of nonword trials, the experiment was again analysed as a two factor design, with type of vowel grapheme and type of case mixing as independent variables.Trimming procedures for the analysis by participants and the analysis by stimuli were the same as for the word data.Mean response times and error rates in the by-participant analysis are shown in Table 1.
There was a significant main effect of the type of vowel grapheme in the analysis by participants, F1(1,19)=27.34,MSE=3308, p<0.01), but not in the analysis by stimuli, with faster responses to nonwords with single-letter vowel graphemes than to nonwords containing multi-letter vowel graphemes.Neither the main effect of type of case mixing nor the two-way interaction was significant.
In an analysis of variance of error rates the only significant effect was a two-way interaction between the type of vowel grapheme and type of case mixing in the analysis by participants, F1(1,19)=5.49,MSE=0.003,p<0.05.

Discussion
The observed interaction between the type of vowel and the pattern of case mixing confirms our initial hypothesis.For words containing MLG vowels, case mixing had a greater impact on recognition latencies when the vowel grapheme was disrupted (e.g.sTeAk), than when it was left visually intact (e.g.stEAk).This effect cannot be attributed to differences in the number of case alternations within the letter string as there was no corresponding effect for control words containing single letter graphemes with similar patterns of case mixing (e.g.gRuNt vs. grUNt).This suggests that preservation of MLGs in a mixed-case letter string through appropriate grouping of letters aids word recognition in a lexical decision task.
Unlike in previous studies (e.g.Rastle & Coltheart, 1998;Rey et al. 1998Rey et al. , 2000)), the case mixing manipulation used in the present experiment has led to the reversal of the previously reported effect, with words containing MLGs being recognised faster on average compared to the words containing single-letter graphemes.This seems to be a consequence of the fact that data for the MLG conditions include trials where the MLG preserving case alteration pattern leads to fast response times.
Our nonword data replicates the "whammy" effect observed by Rastle & Coltheart (1998), only this time in a lexical decision task.However, the differential effect of the case mixing pattern on items with single and multi-letter graphemes was not observed in nonwords.One possible explanation could be found in the nature of the task.According to the DRC model (Coltheart, Rastle, Perry, Langdon, & Ziegler, 2001), making "yes" and "no" responses in a lexical decision task is based on different procedures.A "Yes" decision is determined by the activation level of an entry in the orthographic lexicon, or the sum of the activation of all of the entries in the orthographic lexicon ("fast-guess" mechanism).On the other hand, a "no" decision is made when a sufficient number of processing cycles has elapsed and a "yes" decision has not yet been made.Therefore, it is to be expected that factors which seem to influence a decision based on the level of activation in the orthographic lexicon will be of less importance for the decisions based on the time-based criteria.
EXPERIMENT 2 Experiment 1 demonstrated that recognition is faster for mixed-case words in which the structure of MLG vowels is preserved.A further experiment was designed to replicate this result with a new set of stimuli, and to establish whether this finding extends to graphemes in which the critical MLG represents a consonantal grapheme rather than a medial vowel (e.g.sh, ck).k k

Method
Participants: Twenty students participated in the experiment as part of a course requirement.All participants were native speakers of English, with normal or corrected-to-normal vision.
Stimuli and design: The experiment followed a three-factor, repeated measures design, with lexicality of items (word vs. nonword), type of grapheme (vowel vs. consonant) and type of case mixing ("disruptive pattern" which disrupts MLGs vs. "non-disruptive pattern" which leaves MLGs intact) manipulated as independent variables.
Forty monosyllabic, five-letter words, matched on frequency were selected for this experiment.Twenty of these contained MLG consonants (e.g.graph a ) and twenty contained a MLG vowel (e.g.tread).Half of words from each category were presented with a pattern of case mixing which disrupted MLGs (e.g.tRaIT, smASh), while for the remainder, MLGs were left intact (e.g.trAIt, sMaSH).A set of 40 nonwords were generated by changing the identity H H of the were presented with the same patterns of as corresponding words.Patterns of mixing were counterbalanced across words and nonwords during the experiment.All items, together with their response latencies and error rates, are presented in Appendix 2. Stimuli and the matched nonwords were presented in a randomised sequence during the experiment.An additional set of 24 monosyllabic words and nonwords were used for practice trials.Practice trials had the same pattern of case mixing as experimental trials.
Procedure: The procedure used in this experiment was the same as in the baseline study, with 24 practice trials followed by two blocks of 40 experimental trials, separated by a short rest interval.

Results
Analysis of word trials.Two factors were analysed in the analysis of word trials: type of MLG and type of case mixing.The procedure for trimming the RT data was the same as in the previous experiments.In addition to trimming data points that were two standard deviations above or below the mean, items associated with a high proportion of incorrect responses (30% or more) were also excluded from the analysis.There were four such items; two were words with two-letter vowel graphemes, and the other two contained two-letter consonant graphemes.Mean response times and error rates for word and nonword in the byparticipant analysis are given in Table 2.In an analysis of variance by participants (F1), type of vowel grapheme and type of case mixing were treated as within group factors.In the corresponding analysis of variance by stimuli (F2), type of MLG was treated as a between group factor, and type of case mixing as a within group.Both main effects were significant in this analysis.Responses to words containing MLG vowels were significantly slower than responses to words containing MLG consonants, F1(1,19)=7.04,MSE=3232, p<0.01;F2(1,34)=5.46,MSE=2186, p<.05.Responses were also slower to words in which MLGs were disrupted by case mixing, when compared with words in which MLGs remained intact, F1(1,19)= 4.95, MSE=2233, p<0.04;F2(1,34)=4.37,MSE=6091, p<0.05.However, this time there was no significant interaction between grapheme type and case mixing pattern.
In an analysis of error rates, the only significant effect was the main effect of the type of case mixing in the analysis by participants, F1(1,19)=16,37, MSE=0.003,p<0.01, the main effect of type of grapheme and the interaction were not significant in either of the analyses.
Analysis of nonword trials.In the analysis of nonword trials, the experiment was again analysed as a two factor design, with type of MLG and type of case mixing as independent variables.Trimming procedures for analysis by participants and the analysis by stimuli were the same as for the word data, and the mean response times and error rates are shown in Table 2.
The main effect of the type of MLG was significant only in the analysis by participants, F1(1,19)=10.85,MSE=4603, p<0.01, with nonwords containing multi-letter vowel graphemes being slower than nonwords with multi-letter consonant graphemes.The main effect of type of case mixing and the twoway interaction were not significant in either of the analyses.In the analysis of variance by participants and by stimuli for error rates, neither of the main effects nor the interaction was significant.

Discussion
The significant effect of case mixing replicates the main finding of Experiment 1; words presented in mixed case are recognised more quickly when MLGs are preserved.The present experiment also extends this finding to words containing MLG consonants.However, words with consonant MLGs are processed faster than words with vowel MLGs even when they are disrupted by case mixing.There are several possible explanations for this finding.One explanation is that the spelling and pronunciation of vowels is highly inconsistent: almost every vowel sound can be spelled in more than one way, and almost every vowel spelling could be pronounced in more than one way.On the other hand, the spelling to sound correspondence for multi-letter consonants (and consonants in general) is highly consistent.Therefore, it is possible that what we have observed here is a consistency effect (i.e.Jared, 1997;Stone, Vanhoy, and Van Orden, 1997;Ziegler, Montant, & Jacobs, 1997) at the level of individual grapheme-phoneme correspondences (Zorzi, 2000).Alternatively, a possible factor contributing to faster recognition of words with multi-letter consonant graphemes could be their serial position.In the stimuli used in the present experiment, the MLG consonant was always the final consonant of the word (occupying the last two letter positions), while MLG vowels were always placed in the middle of the word.According to the DRC model (Coltheart & Rastle, 1994;Coltheart, et al., 2001;Rastle & Coltheart, 1999) assembly of phonology operates serially from left to right.Their research has demonstrated that, when reading aloud, the presence of a letter with irregular spelling-to-sound correspondence at the beginning of the letter string leads to slower response times compared with the words in which the "point of irregularity" occurs later in the letter string.What we have observed here could be an operation of the same procedure, but this time observed in a lexical decision task.
Again, there was no effect of the type of case mixing for nonwords, indicating that rejection of nonwords in the lexical decision task is not facilitated by the "grapheme preserving" segmentation of the letter string when the nonwords are not pseudohomophones.

GENERAL DISCUSSION
The present experiments confirm that case mixing can have different effects depending on the type of orthographic unit that is broken up by the manipulation.The disruptive effect of case mixing on recognition times is significantly reduced if orthographic units (i.e.graphemes), which are important for the nonlexical processing of a letter string, are preserved.This finding was replicated in both experiments, and was found with both vowel and consonant multi-letter graphemes.To our knowledge this is the first time that these effects have been reported in the literature.
It is clear that this effect cannot be attributed to purely perceptual factors.If the difference in response times for two different patterns of case mixing was attributable to the degree of visual disruption associated with the given case mixing pattern, we would have expected to find a significant difference between the two case mixing patterns in words with single-letter graphemes.However, in Experiment 1, no such difference was observed.Responses were faster only when letter groupings that were preserved corresponded to MLGs.In other words, the variable that had a critical effect was the same-case grouping of letters within a MLG rather than a number of case alternations within a letter string.
These experiments were not explicitly designed to identify the mechanisms responsible for the case mixing effect.Nevertheless, it is possible to say that the results presented here provide further support for the hypothesis that one of the most important factors in producing the case mixing effect is inappropriate letter grouping (Mayall et al., 1997), as visual grouping of letter clusters that corresponded to a grapheme had a beneficial effect on word recognition times.
Another interesting result observed in the present study was the difference in response times between words that contained MLG vowels and those that contained MLG consonants.Regardless of the case mixing pattern it took longer to process words with MLG vowels than words with MLG consonants.As noted in the discussion of Experiment 2, this effect could be seen as a form of the consistency effect.Vowel graphemes are more inconsistent than consonant graphemes in terms of spelling-to-sound correspondence.Therefore, what we are observing here may be caused by the grapheme-phoneme consistency effect (Zorzi, 2000), where the processing of a letter string which contains inconsistent elements will take longer.Although consistency effects are typically found in reading aloud tasks (e.g.Jarred, 1997), some authors have reported them in lexical decision tasks as well (e.g.Stone et al., 1997).
The main effect of the type of MLG under case mixed conditions may also reflect left-to-right processing of the letter string by the nonlexical route, as suggested by the DRC model (Coltheart et al., 2001), given that vowel MLGs in our experiment occupied an earlier position in the letter string compared to the consonant MLGs.However, this effect may not be easily captured by the current version of the DRC model (Coltheart et al., 2001), as it posits that activation from the nonlexical route in a visual lexical decision task is slow to arise and has little impact on making "yes" decisions.
On the other hand, the overall case mixing effect should be relatively easy to simulate by the DRC model.According to Mayall et al., (1997), disruption of trans-letter features is, together with inappropriate grouping, the main mechanism behind the case mixing effect.In the DRC model, the initial stages of word recognition involve activation of the letter features which then activate the word's letter units.This activation from the letter units is then passed on to the word units in the orthographic lexicon (Coltheart et al., 2001).If case mixing disrupts trans-letter features it may slow down the rise in activation of letter feature units, which will in turn delay activation of the letter units, and consequently the word units in the orthographic lexicon.This would slow down "yes" responses in a lexical decision task.Nevertheless, this account of the case mixing effect in the lexical decision task will probably fail to predict the observed advantage for words where MLG is preserved under the case mixed condition unless we assume that case mixing slows down the activation of the word units in the orthographic lexicon more than it slows down processing via the nonlexical route.This would allow for activation from the nonlexical route to make an impact on "yes" decisions in the lexical decision task.Whether this is possible is difficult to say without running a simulation which is beyond the scope of the present paper.
Even though results from these case mixing experiments cannot be directly generalised to the reading of words presented in the same case, they can give us some information about the relative importance of different visual units in reading and the early stages of visual word recognition.The present results confirm that graphemes are units that play an important role in visual word recognition, and that manipulation of their presentation by case mixing will have a significant effect on response latencies to words in a lexical decision task.As such these findings need to be taken into account by the models of visual word recognition.

Table 1 .
Mean response times (in ms) and proportion of errors in mixed case and baseline conditions for Experiment 1

Table 2 .
Mean response times (in ms) and proportion of errors in mixed case and baseline conditions for Experiment 2