Download as pdf or txt
Download as pdf or txt
You are on page 1of 30

This article was downloaded by: [University of St Andrews] On: 24 April 2012, At: 04:22 Publisher: Psychology Press

Informa Ltd Registered in England and Wales Registered Number: 1072954 Registered office: Mortimer House, 37-41 Mortimer Street, London W1T 3JH, UK

The Quarterly Journal of Experimental Psychology Section A: Human Experimental Psychology


Publication details, including instructions for authors and subscription information: http://www.tandfonline.com/loi/pqja20

Notational modulation of the SNARC and the MARC (linguistic markedness of response codes) effect
HansChristoph Nuerk , Wiebke Iversen & Klaus Willmes
a b a b a

University Hospital of the RWTH Aachen, Aachen, Germany University of Cologne, Cologne, Germany

Available online: 13 May 2010

To cite this article: HansChristoph Nuerk, Wiebke Iversen & Klaus Willmes (2004): Notational modulation of the SNARC and the MARC (linguistic markedness of response codes) effect , The Quarterly Journal of Experimental Psychology Section A: Human Experimental Psychology, 57:5, 835-863 To link to this article: http://dx.doi.org/10.1080/02724980343000512

PLEASE SCROLL DOWN FOR ARTICLE Full terms and conditions of use: http://www.tandfonline.com/page/terms-and-conditions This article may be used for research, teaching, and private study purposes. Any substantial or systematic reproduction, redistribution, reselling, loan, sub-licensing, systematic supply, or distribution in any form to anyone is expressly forbidden. The publisher does not give any warranty express or implied or make any representation that the contents will be complete or accurate or up to date. The accuracy of any instructions, formulae, and drug doses should be independently verified with primary sources. The publisher shall not be liable for any loss, actions, claims, proceedings, demand, or costs or damages whatsoever or howsoever caused arising directly or indirectly in connection with or arising out of the use of this material.

Q2166QJEP(A)05701 / Jun 1, 04 (Tue)/ [29 pages 3 Tables 5 Figures 2 Footnotes 0 Appendices] . Centre single caption cf. [no comma] Disk edited WTG.
THE QUARTERLY JOURNAL OF EXPERIMENTAL PSYCHOLOGY, 2004, 57A (5), 835863

Notational modulation of the SNARC and the MARC (linguistic markedness of response codes) effect
Hans-Christoph Nuerk
University Hospital of the RWTH Aachen, Aachen, Germany

Downloaded by [University of St Andrews] at 04:22 24 April 2012

Wiebke Iversen
University of Cologne, Cologne, Germany

Klaus Willmes
University Hospital of the RWTH Aachen, Aachen, Germany
Number magnitude and number parity representation are fundamental number representations. However, the representation of parity is much less understood than that of magnitude: Therefore, we investigated it by examining the (new) Linguistic Markedness of Response Codes (MARC) effect: Responses are facilitated if stimuli and response codes both have the same (congruent) linguistic markedness (evenright, oddleft) while incongruent conditions (evenleft, odd right) lead to interference. We examined systematically the MARC (for parity) and the Spatial Numerical Association of Response Codes (SNARC; for magnitude) effect for different number notations (positive Arabic, negative Arabic, number words) and with different methods of data analysis. In a parity judgement task, the SNARC effect indicating a magnitude representation was replicated for all notations except for negative numerals. The MARC effect was found for number words in all analyses, but less consistently for the other notations. In contrast, a correlational analysis of the reaction time (RT) data, as suggested by Sternberg (1969) using a nonmetric multidimensional scaling (MDS) procedure, produced a clear association of parity and response code for all notations (MARC effect), but little evidence of the SNARC effect. We discuss the extent to which these notation-specific MARC and SNARC effects constrain current models of number processing and elaborate on the possible functional locus of the MARC effect.

Correspondence should be addressed to Hans-Christoph Nuerk or Klaus Willmes, University Hospital of the RWTH Aachen, Neurology Section Neuropsychology, Pauwelsstr. 30, D 52057 Aachen, Germany. Email: hcnuerk@ukaachen.de or willmes@neuropsych.rwth-aachen.de This research was supported by a DFG (German Research Foundation) grant to Klaus Willmes supporting HansChristoph Nuerk. We would like to thank Stuart Fellows for checking the English grammar. This article is an extended and a more elaborated version of a paper presented by Willmes and Iversen at the Spring Annual Meeting of the British Neuropsychological Society in London on April 56, 1995. 2004 The Experimental Psychology Society http://www.tandf.co.uk/journals/pp/02724987.html DOI:10.1080/02724980343000512

836

NUERK, IVERSEN, WILLMES

Downloaded by [University of St Andrews] at 04:22 24 April 2012

Parity and magnitude are the two most important features/properties of numbers when judgements about the similarity of numbers have to be made (Miller & Gelman, 1983; Miller & Stigler, 1987, 1991; Shepard, Kilpatric, & Cunningham, 1975). In these studies, nonmetric multidimensional similarity scaling methods revealed that the spatial representations obtained were all similar in that numbers were grouped according to magnitude and parity. Miller and Gelman (1983) found that number magnitude was the most salient feature for kindergarten children, whereas parity became more important for school children from Grade 6 on. Adults gave almost identical weight to both magnitude and parity. These similarity judgement studies contrast remarkably with reaction time (RT) studies yielding magnitude effects in tasks, in which magnitude was irrelevant for performance (Berch, Foley, Hill, & McDonough Ryan, 1999, for children from Grade 2 on; for adults; Brysbaert, 1995; Dehaene & Akhavein, 1995; Dehaene, Dupoux, & Mehler, 1990; Fias, Brysbaert, Geypens, & dYdewalle, 1996). In contrast to magnitude, however, parity effects seemed to occur inconsistently, even in parity judgement tasks (e.g., Dehaene, Bossini, & Giraux, 1993, vs. Hines, 1990). This introduction is intended to provide a short survey of magnitude and parity effects. In particular, it will be examined why huge parity effects have been found in some experiments, but null effects in others. A possible account for this divergence will be presented.

Magnitude effects and the SNARC effect


In RT studies, number magnitude seems to lead to the most reliable and stable effects in a variety of tasks. Three effects are most often quoted as evidence for an internal magnitude representation: the distance effect (Dehaene, 1989; Dehaene et al., 1990; Hinrichs, Yurko, & Hu, 1981; Moyer & Landauer, 1967; for an overview, see Butterworth, 1999); the problem size effect (Ashcraft, 1992; Brysbaert, 1995; Buckley & Gillman, 1974; Foltz, Poltrock, & Potts, 1984); and the Spatial Numerical Association of Response Codes (SNARC) effect (Dehaene et al., 1993). The distance effect in number processing describes the finding that the time to compare the magnitude of numbers decreases with the increase in numerical difference between them. It can be accounted for by the assumption of an analogue magnitude representation for numbers, which serves as a medium for the comparison process on a mental number line (Restle, 1970). The problem size effect denotes that performance in a variety of number or calculation tasks usually worsens for relatively larger numbers. The most important magnitude effect for this paper is the SNARC effect (Dehaene et al., 1993; Fias, 2001; Fias et al., 1996). The SNARC effect consists of a systematic interaction between response side and number magnitude: Small numbers in the range of numbers presented are responded to faster with the left-hand key and large numbers faster with the right-hand key. Dehaene and co-workers assumed that during parity judgements an analogue magnitude representation on a left-to-right-oriented mental number line is invoked automatically. This oriented mental number line is assumed to be associated with the response codes (left-hand and right-hand keys) thereby producing the SNARC effect. The SNARC effect is most pronounced for onedigit Arabic numerals and less so for number words in the range of 019. This pattern of results may indicate a stronger relationship between Arabic code and magnitude representation than between verbal code and magnitude representation (triple code model of Dehaene, 1992; Dehaene & Cohen, 1995, 1997). Fias and colleagues (Fias, 2001, Fias et al., 1996)

NOTATIONAL MODULATION OF SNARC AND MARC EFFECT

837

Downloaded by [University of St Andrews] at 04:22 24 April 2012

showed an interesting interaction between notation and task for the SNARC effect. While for Arabic numerals the SNARC effect could be obtained for parity judgements and phoneme detection, for number words a SNARC effect was only observed for parity judgement. This result led Fias (2001) to postulate an additional asemantic route for processing number words in asemantic tasks like phoneme detection for number words. The SNARC effect could be generalized to two-digit numbers (Dehaene et al., 1993). Thus, an analogue left-to-right-oriented compressed number line can be assumed for all numbers from 0 to 99 (Brysbaert, 1995; Reynvoet & Brysbaert, 1999). However, for negative numbers, it remains unclear what the orientation of the mental number line, as indexed by the SNARC effect, might be. Recently, it has been shown that the direction of the SNARC effect can be reversed with an appropriate visual anchor. Bchthold, Baumller, and Brugger (1998) found a reversal when they presented a clock and asked whether a number presented in the centre of the clock was indicating a later or an earlier time than 6 oclock. Since a reversal of the SNARC effect is possible, it is an open question whether it occurs for negative numbers because the absolute value of negative numbers increases from right to left. Vice versa, one could also argue that the mental number line simply proceeds from left to right as in any graph including positive and negative numerals. In this case, a nonreversed SNARC effect should be obtained for negative numerals.

Parity effects: Their assessment and their account in reaction time tasks
In contrast to magnitude effects, parity effects are more controversial. When parity effects are observed, they tend to be stronger for number words than for Arabic numerals. While Hines (1990) observed an odd effectthat is, odd numbers being responded to more slowly than even numbersDehaene et al. (1993) failed to find main odd effects. However, these differences can be explained by a closer look at the stimuli, the notation, and the task used in these experiments. Let us first consider the results reported by Hines (1990). In simple parity judgement tasks, he obtained an odd effect for RT for number words (Experiment 5), but not for Arabic numerals (Experiment 2). For Arabic numerals, there was only an odd effect for errors. Dehaene et al.s (1993) failure to find a main odd effect for RT in Arabic numerals in their Experiment 1 is thus a replication of Hines finding. In their Experiment 9, Dehaene and coworkers obtained an odd effect when number words and Arabic numerals were analysed together for the range 09, thus again replicating the finding of Hines. In contrast to Hines, however, they did not find an odd effect for number words in their Experiment 8 (nor in their Experiment 9 for the range 019). The reason for these inconsistent results seems to be the inclusion or exclusion of zero. While Hines employed stimuli from 2 to 9 to examine the odd effect, Dehaene et al. (1993) used stimuli from 0 to 9. The 0 in Dehaene et al.s Experiment 9 was about 60 ms slower than the average for other even numbers from 2 to 8 and still about 40 ms slower in Experiment 8 (see Figures 13 and 15). Given that Hines (1990) reported an odd effect of 22 ms in number words for the range 2 to 9, it is quite reasonable to argue that the inclusion of zero practically levels out any main odd effect: The inclusion of zero (1) slows down the average response to even numbers and (2) greatly increases the mean square errors contributing to the F statistics in a repeated measures analysis of variance (ANOVA). This

838

NUERK, IVERSEN, WILLMES

argument is further confirmed by Dehaene et al.s post hoc comparisons: Responses to zero were slower than those to every other even number (except for number 6 in some experiments). Most researchers seem to agree that zero is not a typical even number and should not be investigated as part of the mental number line (see Brysbaert, 1995; Fias, 2001; see also Armstrong, Gleitman, & Gleitman, 1983, who showed that not all numbers are subjectively equally odd or even). Given that untrained participants are sometimes unsure about the status of zero, rejection of the odd effect seems premature when based on a null effect that is only obtained with inclusion of zero. Sometimes Dehaene et al. (1993) nevertheless found an odd effect because they also included the number 1, which is also responded to more slowly than other odd numbers. However, inclusion of the number 1 does not compensate for inclusion of zero because (1) the RT difference between zero and other even numbers is usually much larger than the RT difference between 1 and other odd numbers and because (2) the inclusion of zero increases the mean square error in the denominator of the F statistic even further. Whether or not zero is similar to other even numbers could be assessed via a correlational analysis using multidimensional scaling (which will be done in this study). If it is similar, it should be close in similarity space. Otherwise, it should be far apart from the other even numbers. In sum, a main odd effect seems to occur at least for one-digit number words when zero is not included. Thus, the odd effect seems to be dependent on the range of the stimuli and the notation in the parity judgement task. However, there are huge alterations of the odd effect. In Hines (1990), Experiment 1, participants had to decide whether the numbers had the same (unmarked) or a different (marked) parity. In this experiment, the odd effect was very large compared to the 22 ms in the simple parity judgement task above: about 150200 ms for young participants and about 300 400 ms for older participants. So far, there is little theoretical background to account for the odd effect being more pronounced in Hines (1990) samedifferent tasks than in the parity judgement task, as well as being more pronounced for verbal notation. Hines explained the odd effect by pointing out that the adjective even is linguistically non-marked as opposed to the marked adjective odd. Nonmarked linguistic entities are assumed to be retrieved more quickly. However, this explanation does in our opinion not suffice to explain why the odd effect is so much larger in Hines task than in other parity judgement tasks. As a possible hypothesis, we wish to put forward and to test in this paper a more general version of the linguistic markedness hypothesis. In most languages there are pairs of complementary adjectives like odd and even. The nonmarked adjective is determined by prefixing both adjectives with the syllable un or in or other prefixes negating the original attribute. Only the nonmarked exemplar (un-even) can be negated this way (Zimmer, 1964). Apart from this formal type of markedness, Lyons (1969) mentions semantic (rightleft, samedifferent) and distributive markedness, which is related to word frequency (lion-lioness). Markedness has been shown to affect response times in other domains. Sherman (1973, 1976) has shown that marked adjectives lead to longer sentence comprehension times. The more general version of the linguistic markedness hypothesis would not only account for the main odd effect, but also for the huge enhancement of the odd effect in Hines (1990) samedifferent parity judgement experiment in the following way: When people have to

Downloaded by [University of St Andrews] at 04:22 24 April 2012

NOTATIONAL MODULATION OF SNARC AND MARC EFFECT

839

respond same (parity) to two even numbers, they have to respond with an unmarked response (same) to an unmarked stimulus attribute (even). Stimulus and response are thus congruent with respect to linguistic markedness. However, if they have to respond with the unmarked response same to a marked stimuli attributes (odd), stimulus and response are incongruent with respect to linguistic markedness. The markedness-incongruent odd same condition was much slower than the markedness-congruent evensame condition. This difference was 520 times as large as the usual odd effect. We thus hypothesize that it is not only the markedness of the stimulus that is important (as suggested by Hines), but also the congruency of the markedness attributes of stimuli and responses.

Downloaded by [University of St Andrews] at 04:22 24 April 2012

EXPERIMENT 1 Average-based ANOVA and regression analyses


If this linguistic markedness hypothesis is true, the predictions for a parity judgement task, in which the hand-to-response relation is manipulated within participants, are straightforward: the adjectives right and even are linguistically nonmarked (Zimmer, 1964). On the contrary, the adjectives left and odd are linguistically marked. If the linguistic markedness hypothesis holds, one should observe interference if unmarked responses are associated with marked response keys and vice versa (markedness incongruent condition: oddright, evenleft), while one should observe facilitation if the markedness association between stimulus and response is congruent (evenright, oddleft). We would like to call this hypothesized effect the Markedness Association of Response Codes (MARC) effect (cf. Willmes & Iversen, 1995; see also Berch et al., 1999; Reynvoet & Brysbaert, 1999) in analogy to the SNARC effect: If the MARC effect can be demonstrated, then the two most important semantic features (magnitude and parity) of numbers can be associated with certain response codes. We tested this hypothesis with a typical oddeven judgement task in this study for different notations in a within-participant design. Parity effects have been shown to differ between notations. Mostly, they tend to be stronger for verbal notation (number words). Linguistic markedness may also have a greater effect for number words than for Arabic numerals, because of the verballinguistic attribute of the concept (cf. Hines, 1990, for early indications of such a notation-specific effect and an early suggestion that markedness may exert stronger influences on verbal notation). If this hypothesis about notational modulation were true, the hypothesized MARC effect (evenright, odd left being faster than oddright, evenleft) should be observed particularly for those notations that most likely activate verballinguistic concepts. Therefore, we expect a strong MARC effect for verbal notation (number words) and less for positive and negative Arabic notation. If linguistic markedness is an important dimension for the similarity of responses to Arabic numerals, congruity between the markedness of number and response should be structuring multidimensional similarity space as well. If the MARC effect only rests on responses to a few numbers, linguistic markedness should not play such a role in multidimensional similarity space. With regard to current models of number processing, to our knowledge there is no model that has yet linked number parity to linguistic concepts and properties. Neither the possible stimulusresponse markedness congruency nor the possible notation specificity of possible MARC effects have yet been incorporated in any model.

840

NUERK, IVERSEN, WILLMES

Downloaded by [University of St Andrews] at 04:22 24 April 2012

Besides verbal and positive Arabic notation, the goal of this study is to examine the SNARC effect and its underlying representation more generally by also investigating negative Arabic numbers in a within-participant design. As laid out above, the SNARC effect is interpreted spatially as evidence for a left-to-right-oriented mental number line. Larger numbers (on the right side) of the number line are responded to faster with the right hand in space. However, it is interesting to question how the SNARC effect generalizes to negative numbers. Dehaene and colleagues (Dehaene et al., 1993) postulated that the SNARC effect is due to reading habits, as it was not found for Iranian participants living in France only for a short time. An alternative account would be that the number line is oriented form left to right simply because this is the way number lines are depicted in arithmetic graphs in Western culture. If arithmetic graphs are so salient that they are the source of the SNARC effect, then the SNARC effect for negative numerals should be reversed, in that negative numerals with higher absolute values are farther left on the number line in mathematical function graphs. If reading habits are the major source of the SNARC effect then there is no obvious reason why the minus sign before a number should change the direction of the SNARC effect. To examine this question, positive and negative Arabic integers were directly compared in a within-participant design.

Method
Participants
A total of 32 German participants, 11 female and 21 male, aged between 23 and 30 years (median 26 years) were tested individually. They were students from different faculties of the Technical University of Aachen. A total of 31 were right-handed, and 1 was left-handed. The left-handed participant was included since his performance pattern was not different from that of the right-handed participants.

Stimuli
Participants were presented with positive Arabic digits in the range 0 to 9, negative Arabic digits in the range 9 to 0, German number words from eins (one) to neun (nine), as well as Roman numerals from I to IX (which are not reported here). Each notation was presented in a separate block. In each block a short instruction was given informing the participant about the notation and range of numbers to follow.

Procedure
Participants had to decide whether each number was even or odd by pressing one of two response keys. The experiment was subdivided into two parts. Half of the participants started with the even response assigned to the right-hand key and the odd response assigned to the left-hand key. After responding to the first part of the experiment there was a short break before the second part was started with the same sequence of blocks as that in the first part but with reverse assignment of parity to response key. Each of the two parts was preceded by a training list of 12 items (3 of each notation). In both parts of the experiment, each participant had to respond to four blocks corresponding to the four notations. Eight-block sequences were constructed such that presentations of positive and negative numerals in consecutive blocks were avoided. The target numbers of one block were presented four times in random order. The experiment was controlled by an IBM-compatible PC (4/86-dx/66 Mhz) using the Experimental Run-Time System (ERTS; Beringer, 1993). The targets were presented on a 17-monitor screen

NOTATIONAL MODULATION OF SNARC AND MARC EFFECT

841

using white symbols against a blue background. Responses were recorded by way of two response keys placed before the participant at a distance of 60 cm. At the beginning of each trial, a cross was presented in the centre of the screen for 500 ms. Then the target appeared and remained until response, but maximally 1500 ms. Afterwards the screen remained blank for 1000 ms before the next trial started. The whole experiment lasted for approximately 30 minutes. The number words were given in small print with a maximum height of 10 mm and a maximum length of 28 mm. All other targets also had a height of 10 mm.

Results
For comparability between the notations without zero, zero was also excluded in most analyses for positive and negative Arabic numerals. In all analyses, only those 23 participants were considered for whom at least one correct response was available for every numeral in every presented notation. However, if one performs the analyses with the maximal number of participants available for each single notation, the results do not change substantially. 1. RT analysis ANOVA analyses For all further analyses, the median RT for correct answers was computed for each target, each side of response, and each participant. It was thus based on a maximum of four correct responses. For negative numerals the labels of the target magnitude factor represented the absolute values of the numbers presented in the analysis (for descriptive values, see Tables 1 and 2). Positive Arabic numerals. We obtained significant main effects of magnitude and parity, F(3, 66) = 6.58 and F(1, 22) = 29.80, respectively, both p < .001. For parity, the odd effect (Hines, 1990) was replicated: Responses to even number responses were faster than to those odd ones. For magnitude, a quadratic trend reached significance, F(1, 22) = 8.58, p < .01: Responses to large and small magnitudes tended to be faster. Response hand, in contrast, failed to reach significance (F < 1). Two 2-way interactions were (marginally) significant: Parity Magnitude, F(3, 66) = 4.14, p < .01 and Magnitude Response hand, F(3, 66) = 2.48, p = .07, indicating a SNARC effect, see Figure 1: The Parity Hand interaction (MARC effect) failed to reach significance, F(1, 22) = 1.93, p = .17. However, when the hypothesized MARC effect is tested directly (MARC-congruent against MARC-incongruent conditions), a marginally significant effect remains in that responses in congruent trials were 16 ms faster than those in incongruent trials, t(22) = 1.41, p = .09, see Figure 2; Table 1. Finally, there was a three-way interaction between magnitude, parity, and response hand for positive Arabic numerals, F(3, 66) = 3.01, p = .05, GreenhouseGeisser adjusted, which seemed to indicate a greater SNARC effect for odd numbers. Negative Arabic numerals. We only obtained one significant main effect: the odd effect, F(1, 22) = 7.62, p < .001. No other main effect or interaction reached significance (all p .10). Number words. For number words, the results were quite the opposite of the results for the positive Arabic notation. All parity effects and interactions were clearly significant, while

Downloaded by [University of St Andrews] at 04:22 24 April 2012

TABLE 1 Overview of the most important average median RT effects Arabic Overall Odd Even Odd effect Left hand Right hand Hand effect MARC-compatible MARC-incompatible MARC effect MARC (Mult. Reg. ) SNARC (Lin. Reg. ) SNARC (Mult. Reg. ) 509 529 489 40 509 508 1 501 517 16 12 10.06 8.89 Negative Arabic 518 529 508 21 518 519 1 517 520 3 0 5.70 5.68 Number words 551 566 535 31 552 549 3 532 569 37 34 8.35 7.53

Downloaded by [University of St Andrews] at 04:22 24 April 2012

Note: For detailed RT differences regarding the SNARC effect see Table 2. SNARC (Lin. Reg. ) denotes the mean SNARC effect as it is usually determined in a simple linear regression. SNARC (Mult. Reg. ) and MARC (Mult. Reg.) denote the SNARC effect and the MARC effect, respectively, when magnitude and parity are both included in a multiple regression over left handright hand differences. Although both effects decrease slightly in the multiple regression, the overall pattern between notations is fairly stable.

TABLE 2 Average RT values for each number in each notation and for each response hand Positive Arabic Num. abs left right diff 1 2 3 4 5 6 7 8 542 478 547 484 537 517 491 477 518 483 531 477 525 545 510 486 565 473 563 490 549 489 471 468 46 9 32 12 23 56 38 18 Negative Arabic abs left right diff 531 499 517 494 541 509 527 531 507 502 525 482 549 507 526 546 555 495 508 506 532 511 528 516 48 7 17 24 17 4 1 29 Number words abs left right diff 589 535 562 522 549 547 566 537 554 537 554 547 539 570 550 569 623 533 569 497 559 524 582 504 69 4 15 50 20 46 33 -64

Note: Abs refers to the overall response time, left and right refer to the left and right hand response times, respectively, and diff refers to the difference between right-hand response and left-hand response computed for analyses of the SNARC effect.

842

Reaction Time

Errors

Downloaded by [University of St Andrews] at 04:22 24 April 2012

Figure 1. SNARC and MARC effects for the RT data (to the left) and errors (to the right): A SNARC effect can be observed when the negative regression slope (see the regression line) explains a large portion of the variance. The MARC effect can be observed by comparison of the dotted line (odd numbers) and the dashed line (even numbers) with each other and with the regression line. For a perfect MARC effect the dotted line for odd numbers should be both above the regression line and above the dashed line for even numbers. Figure 1a (top left): RT positive Arabic numerals. A SNARC effect (slope = 8.89 ms) can be observed, but no significant MARC effect because the differences for the 4 and 7 are on the wrong side of the regression line, thus producing too much error variance. Figure 1b (middle left): RT negative Arabic numerals. Neither a SNARC nor a MARC effect could be observed. Figure 1c (bottom left): RT number words: A SNARC effect (slope = 7.53 ms) and a strong MARC effect (b = 34.07) can be observed. Figure 1d (top right): Errors positive Arabic numerals. Figure 1e (middle right): Errors negative Arabic numerals. Figure 1f (bottom right): Errors number words: For errors, a SNARC effect could be observed for all notations. A significant MARC effect could be seen for number words.

843

844

NUERK, IVERSEN, WILLMES

Downloaded by [University of St Andrews] at 04:22 24 April 2012

Figure 2. MARC effects for RT (top) and errors (bottom) for different notations. A MARC effect is an interaction between parity and hand in this graph. Clear MARC effects were obtained for number words, (marginally) significant effects for positive Arabic numerals, and null effects for negative Arabic numerals.

magnitude seemed to play a lesser role. We observed a highly significant odd effect, F(1, 22) = 19.36, p < .001, and a significant Parity Hand interaction, F(1, 22) = 8.72, p < .01. MARCcongruent trials were clearly faster (37 ms) than MARC-incongruent trials, t(22) = 2.95, p < .01; see Figure 2. Finally, a Magnitude Response Hand interaction was significant, F(3, 66) = 3.01, p < .05, indicating a SNARC effect. Regression analyses The ANOVA may be too conservative because it tests all possible differences between conditions and not only between those specified by the SNARC and MARC hypotheses. Therefore, we tested the SNARC and MARC effect more specifically: In an analysis

NOTATIONAL MODULATION OF SNARC AND MARC EFFECT

845

Downloaded by [University of St Andrews] at 04:22 24 April 2012

suggested by Lorch and Myers (1990) and applied to the SNARC effect by Fias et al. (1996), we performed a multiple regression over the average difference in RT (dRT: rightleft hand) for each number for each individual participant. Then we tested whether the individual nonstandardized regression weights for magnitude and parity differed from zero across participants. A significant SNARC effect was obtained for all notations, all t(22) > 2.30, all p < .02, except for negative Arabic numerals: only marginally significant, t(22) = 1.51, p = .07. A MARC effect was again found for number words, t(22) = 2.69, p < .01. For Arabic integers, once again, no MARC effect could be observed, both t(22) < 1.10, p > .14. The comparison of the regression slopes of the SNARC effect revealed some surprising results. All comparisons between the slopes from different notations were not significant. In particular, the slope for negative Arabic numerals was also not significantly flatter than that for positive Arabic numerals, t(22) = 0.545, p = .30. Thus, previous differences in significance between positive and negative Arabic numerals must be interpreted with caution because the two notations do not significantly differ in the direct comparison of their SNARC effect. In contrast, the MARC beta weights were larger for number words than for negative Arabic numerals, t(22) = 1.85, p < .05, and marginally larger than those for positive Arabic numerals, t(22) = 1.47, p = .08. Thus, the greater MARC effect for verbal notation tended to be confirmed in this analysis. 2. Error analysis Because of the rather high proportion of errors, we also performed the same analyses for errors to examine whether SNARC and MARC effects are due to speedaccuracy trade-off effects.1 Contrary to our expectations, we found significant effects in the error analyses. They were almost always in the same direction as in the RT analysis: Slower conditions tended to be more error prone. Since to our knowledge SNARC effects have rarely been reported for error data (and MARC effects not at all), we report the error SNARC and MARC effects butdue to space restrictionswe do not discuss other effects in detail. ANOVA analyses Positive Arabic numerals. No main effects were observed; however, all two-way interactionsand in particular the SNARC- and MARC-type interactionsreached significance: Parity Magnitude, F(3, 66) = 8.89, p < .001; Magnitude Response Hand, F(3, 66) = 3.48, p < .05; indicating a SNARC effect; and Parity Hand interaction, MARC effect, F(1, 22) = 5.42, p < .05. With regard to the Magnitude Hand Interaction, a clear SNARC effect was observed, F(1, 22) = 7.20, p = .01: For the right hand, error rates decreased with increasing magnitude while for the left hand the effect was reversed. The Parity Hand interaction revealed a clear MARC effect in that MARC-compatible trials were more accurate than MARC-incompatible ones, t(22) = 2.33, p = .01. The Parity Magnitude interaction revealed a similar pattern as that in the RT analyses: Finally, there was a marginally significant threeway interaction between magnitude, parity, and response hand for positive Arabic numerals, F(3, 66) = 2.07, p = .05.
1

We thank Robert Logie for suggesting this analysis.

846

NUERK, IVERSEN, WILLMES

Negative Arabic numerals. The three-factorial ANOVA over errors for negative Arabic numerals revealed a main effect of parity, F(1, 22) = 6.47, p < .05, indicating a reversed odd effect. Even numbers produced more errors than odd numbers. Most importantly, there was a marginally significant SNARClike Magnitude Hand interaction, which was not found in the RT analysis for negative numerals. As indicated by a significant linear-by-linear contrast, F(1, 22) = 6.48, p < .05, larger numbers tended to produce more errors with the left hand while for smaller numbers the opposite occurred. No other interaction reached significance. Number words. The error analyses revealed interaction effects both of magnitude and of parity in the same direction as that of the RT data. First, however, right-hand trials produced again more errors than left-hand trials, F(1, 22) = 8.44, p < .01. Most importantly, all two-way interactions reached significance. The MARC effect was also found for the error data, F(1, 22) = 8.53, p < .01: MARC-compatible trials were less error prone than MARC-incompatible trials. A SNARC-like Magnitude Hand interaction was observed, F(3, 66) = 3.60, p < .05. For right hand responses, smaller numbers tended to produce more errors, while for left-hand responses, larger numbers tended to produce more errors, F(3, 66) = 10.65, p < .01. Finally, a Magnitude Parity interaction was significant, indicating that the different numbers produced a different number of errors, F(3, 66) = 12.71, p < .001. The 3-way interaction did not reach significance, F < 1. Error ANOVA summary. The only indications of SATOs were main effects of hand (right-hand responses tended to be more error prone in some notations) and a reversed odd effect for negative numerals. SNARC-like interactions were observed in all notations, and MARC effects were observed for number words and positive Arabic numerals. Whenever SNARC- or MARC-like interactions were found, they were always in the same direction as that for the RT data. Regression analyses Many participants were error free in many conditions. Therefore, individual regression analyses do not, in our opinion, make much sense. However, the average error rates can be analysed in a regression analysis because there are enough data points for n = 23 participants to produce sufficient (and more than zero) errors for each condition. In a multiple regression with the predictors number and parity over the average percentage of error difference between right-hand and left-hand responses, we replicated the SNARC effect for positive Arabic numerals, t(5) = 2.46, p < .05, b = 1.75, negative Arabic numerals, t(5) = 3.34, p < .01, b = 1.29, and number words, t(5) = 3.67, p < .01, b = - 1.52. Thus, as in the error ANOVA, a SNARC effect could be observed for all notations, even for negative Arabic numerals for which the SNARC analyses only produced marginally significant effects in the same direction or null effects (see Figure 3). We only found a MARC effect for number words, t(5) = 3.20, p = .01, b = 6.09; there was no significant effect for all other notations. For positive Arabic numerals, for which a MARC effect was obtained in the error ANOVA, the predictor parity failed to reach significance in this analysis, t(5) = 1.32, p = .12, b = 0.37. In sum, we observed a significant SNARC effect for all notations in the error analysis, and a significant MARC effect in the error analyses for number words. For positive Arabic numerals

Downloaded by [University of St Andrews] at 04:22 24 April 2012

NOTATIONAL MODULATION OF SNARC AND MARC EFFECT

847

Downloaded by [University of St Andrews] at 04:22 24 April 2012

Figure 3. Smallest space analysis (SSA) for different notations for the numbers 1 to 8 responded to with the right hand (squares) and the left hand (diamonds). For convenience, MARC-compatible conditions (solid) and MARCincompatible conditions (open) are depicted differently. For all notations parity, magnitude, and response hand did not lead to a separate grouping of different conditions, but only MARC compatibility or incompatibility (see the line that separates every single compatible from every single incompatible condition; no such line could be drawn with respect to parity, magnitude or hand of response). Figure 3a (top left): positive Arabic numerals. Figure 3b (bottom left): negative Arabic numerals. Figure 3c (top right): number words.

there was an indication of a MARC effect in the ANOVA, but not in the regression analysis. Except for the obtained SNARC effect for negative Arabic numerals, the error analyses mirrored those of the RT analyses. In particular, no indication of a speedaccuracy trade-off was observed for any SNARC or MARC effect.

Discussion of average-based ANOVA and regression analyses


The foremost aim of the RT and error analyses was to investigate the linguistic markedness association of response codes (MARC) effect. A MARC effect was consistently observed in all analyses for verbal notation, but not for Arabic integers. Only a trend for positive Arabic integers was obtained when markedness compatibility was directly compared and in the error analyses. These results seem to confirm that linguistic markedness plays a role for parity

848

NUERK, IVERSEN, WILLMES

Downloaded by [University of St Andrews] at 04:22 24 April 2012

decisions. This seems particularly to be the case for number words, indicating that (1) the verbal notation may particularly trigger linguisticsemantic concepts such as markedness and that, vice versa (2) the MARC effect is indeed verballinguistic in nature. In summary, the differences in the MARC and the odd effect between different notations are hard to reconcile with the assumption that parity is retrieved from one single abstract semantic code, which is the same for all notations. With regard to magnitude, we mostly obtained statistical trends of a SNARC effect for negative Arabic numerals, while for positive Arabic integers we observed a SNARC effect in all analyses. These statistical trends, however, should not be totally neglected, because in direct comparisons positive and negative Arabic numerals did not differ significantly with respect to the slope of the regression line. However, when a SNARC trend for negative Arabic numerals was observed, it corresponded to the spatial position of the absolute value of these negative Arabic numerals rather than to the spatial position of their raw (negative) values. Thus, the spatial representation of the mental number line does not simply continue for negative numerals to the left of zero, but it becomes reversed. Smaller negative numerals (with a larger absolute value) tend to be met with faster responses with the right-hand key, but the results are less consistent and conclusive than for positive Arabic integers. Although the spatial representation of the raw negative values in mathematical function graphs may produce some interference in some participants, this does not determine the SNARC effect. Rather, the data are in line with Dehaenes reading habit hypothesis about the SNARC effect (Dehaene et al., 1993).

EXPERIMENT 1 Non-metric multidimensional scaling analyses


While for the RT difference analysis magnitude affected performance in every notation, the effects for parity were less consistent. However, in several studies asking for judgements about the similarity of numbers, nonmetric multidimensional scaling (MDS) methods have revealed that the spatial representations obtained were all similar in that numbers were grouped according to magnitude and parity (Miller & Gelman, 1983; Miller & Stigler, 1987, 1991; Shepard et al., 1975). In particular, Miller and Gelman (1983) found that adults gave almost identical weight to magnitude and parity. Thus, there is a difference between the results of RT experiments and similarity judgements. We wanted to investigate whether this difference is a matter of the task (e.g., speeded parity judgement vs. similarity judgement) or whether it is rather a matter of the analysis method chosen (i.e., investigating mean differences vs. correlational patterns). The idea to analyse the structure of intercorrelations among experimental conditions is not new. Its usefulness is laid out in the last part of the famous additive factor logic article of Sternberg (1969). While the first part concerning the interaction of factors has often guided RT analyses in cognitive psychology (see Sternberg, 1998, for a variety of applications in different areas), the last part about intercorrelations seems to have been considered less often. According to the logic of the additive-factor method (Sternberg, 1969, pp. 308-309) patterns of correlations in addition to interactions among factors reveal which experimental factors influence the same processing stage. The most strongly interacting factors should be most

NOTATIONAL MODULATION OF SNARC AND MARC EFFECT

849

Downloaded by [University of St Andrews] at 04:22 24 April 2012

highly correlated. In number processing research with RT as dependent variable, as far as we know, differences in mean RT among conditions have been analysed exclusively. In the following, we report on nonmetric MDS analyses of patterns of correlations among RT data across all numbers for both response sides within each notation as well as of the correlations between dRT differences (rightleft) for each number in each notation. For these data, a measure of monotone relationship (monotonicity coefficient 2; Borg and Lingoes, 1987, p. 79) was computed between all pairs of stimuli. The resulting correlation matrix was entered into a nonmetric multidimensional similarity scaling procedure (Smallest Space Analysis, SSA-I; Guttman, 1968; Lingoes, 1973). Points that are close together in the SSA configuration represent stimuli that are highly correlated. With regard to a multidimensional similarity scaling procedure, the hypotheses are clear. If similarity in parity and markedness of response is important, odd and even numbers should be grouped together. In particular, for even as well as for odd numbers, all compatible conditions (e.g., 2right, 5left) should be grouped close together, as should all incompatible conditions (e.g., 2left, 5right). However, if the linguistic markedness hypothesis is correct, compatible and incompatible responses should not be very similar and should be located relatively far apart from each other. If magnitude is an important similarity factor, numbers of similar magnitude should be grouped together, while numbers with a large distance in magnitude should be relatively far apart. Thus, if an internal mental number line exerts a strong influence on response similarly in all participants, the multidimensional scaling procedure should group the numbers according to magnitude. Finally, the multidimensional scaling procedure will help us to provide an answer about the status of zero in different notations. If zero is not part of the mental number line (e.g., Brysbaert, 1995) it should be located farther apart from the other numbers. In contrast, if it is part of the number line and if it is just an ordinary even number, it should be located close to other small and to other even numbers. As in the average RT analysis, we have carried out the SSA-I procedure for both absolute RT values and the RT difference between right-hand and left-hand responses. Clearly, with respect to the RT difference, numbers of the same magnitude should again be grouped together (SNARC), and numbers of the same parity should be grouped together (MARC). Again, we hypothesized that the MARC effect should be strongest for the verbal notation.

Results
In this study, nMDS is used to map the number stimuli (different numbers for different response hands) in a space of low dimensionality such that stimuli with high (positive) correlations are mapped close to each other while stimuli with low (negative) correlations are placed far apart. In nonmetric MDS only the pattern of ordinal relations among correlations (larger vs. smaller) is mapped into the corresponding pattern of ordinal relations among distances between stimulus points (Borg & Lingoes, 1987). If no specific a priori hypothesis about the dimensionality of the MDS solution has been formulated, one usually chooses the lowest dimensional solution with acceptable fit as indicated, for example, by the coefficient of alienation (see Borg & Lingoes, 1987, for mathematical definition). A coefficient of alienation for 2D solutions is usually considered adequate in the range of about [.10, .25], particularly, if a theoretically relevant interpretation can be given. If the coefficient of alienation is substantially higher then solutions for higher dimensionality are sought. If the coefficient of alienation

850

NUERK, IVERSEN, WILLMES

is very close to zero, then the solution is very often degeneratefor example, all individual stimulus points are very close together, virtually mapped onto one or two data points of the solution space. In contrast to metric MDS or other metric multivariate dimension-reducing procedures, the dimensions of the solution space themselves are not interpreted. It is the configuration of stimulus points in space that matters (Borg & Lingoes, 1987). The coefficient of alienation as a measure of monotone fit for the spatial representation was acceptably low in the 2-D solution (Arabic numerals, .180; number words, .225; negative numerals, .100; see Figure 4). For all notations the SSA solutions were very similar in that a straight line could be drawn, dividing the space into two regions. On one side of the straight line all points corresponding to the compatible (righteven and leftodd) conditions were located, whereas on the other side all points corresponding to the incompatible (rightodd and lefteven) condition could be found. Concerning number magnitude, no clear-cut pattern could be detected, although in the ANOVAs magnitude had a main effect in all conditions. To investigate the special role of zero in the mental number line, we also computed an SSA including zero for positive and negative Arabic numerals: The points representing zero were well separated from all the other numbers (see Figure 5). Thus, the results of the multidimensional scaling analysis are consistent with the idea that zero is represented differently than other numbers in a parity judgement task. As for the RT analysis, we also explored the results of the multidimensional scaling analysis for the difference in response times with the right and left hand (SNARC and MARC effects). The coefficient of alienation as a measure of monotone fit for the representation was acceptable in two dimensions for positive Arabic integers (.117), while for negative Arabic numerals (.001) and for number words a degenerate solution was obtained (alienation: .001 and .003, respectively). Both degenerate solutions were such that odd trials were mapped virtually onto the point (100, 100) in the MDS space, while the even trials were mapped onto the point (100, 100). While the exact coordinates are meaningless in the MDS, it is very meaningful that now the MARC effect alone determined mapping in the MDS, which was virtually independent of the identity of number and response hand. Thus, these degenerate solutions provide even stronger evidence for the determination by the MARC effect. For positive Arabic integers, the SSA solution exhibited two aspects for rightleft RT difference similarity: A parity aspect (the MARC effect) and a magnitude aspect: While all even numbers were located on one side and all odd numbers on the other (see Figure 5), numbers of greater magnitude were located more in the top part and numbers of smaller magnitude more in the lower part (thus possibly indicating the SNARC effect). But there were two exceptions from a perfect ordering by number magnitude: The number 2 was located between 4 and 6, and the number 7 was located very close to 5 but a little less distant from the number 3 (see Figure 5). Thus, for the similarity in RT rightleft differences, a parity dimension and a magnitude dimension seem to determine how similar RT patterns to different numbers are. For negative Arabic numerals and number words, the degenerate solutions were due to the fact that every single correlation of numbers of the same parity was positive, and almost every single correlation of numbers of different parities was negative (except r = +.04 for the number words seven and eight). Given such a consistent correlation pattern, the SSA simply groups all even and all odd numbers closely together with a maximal distance between the groups. The second dimension played no role in the similarity structure any

Downloaded by [University of St Andrews] at 04:22 24 April 2012

Downloaded by [University of St Andrews] at 04:22 24 April 2012

Figure 4. SSA for positive (top) and negative (bottom) Arabic numerals 0 to 8 responded to with the right hand (squares) and the left hand (diamonds). MARC-compatible conditions (solid) and MARC-incompatible conditions (open) are depicted differently. Again, parity, magnitude, and response hand did not lead to separate grouping of different conditions, but only MARC compatibility or incompatibility. It can also be seen that zero strongly differs from all other numbers regardless of whether it is responded to with the left or the right hand. (See the line that separates zero from the other numbers.)

851

Downloaded by [University of St Andrews] at 04:22 24 April 2012

Figure 5. SSA for differences between right-hand and the left-hand responses for positive (top) and negative (bottom) Arabic numerals, either even (solid diamonds) or odd (open diamonds). It can be seen that the numbers are grouped according to parity. However, this is the only SSA in which magnitude also may play a role (with greater numbers being in the top part rather than in the bottom part) for positive Arabic numerals.

852

NOTATIONAL MODULATION OF SNARC AND MARC EFFECT

853

more. So, if there was any magnitude effect, it could not be observed because the scaling result was completely determined by parity (respectively the MARC effect) for those two notations. In sum, three results are important: The interaction of parity and response hand (MARC congruity) predominantly determined the SSA similarity structure for all notations, while parity and magnitude themselves did not seem to shape similarity structure much. Second, for the analysis of similarity of rightleft-hand RT differences, the parity dimension (i.e., in its MARC interaction with hand) again determined similarity for all notations, while magnitude only determined similarity for positive Arabic integers. Finally, zero could be clearly separated from all other positive and negative Arabic integers in the SSA analysis.

Downloaded by [University of St Andrews] at 04:22 24 April 2012

Discussion of nonmetric multidimensional scaling analyses


Summary and interpretation of nonmetric MDS results. In line with Sternberg (1969), we can conclude on the basis of the SSA that the markedness of response hand and the markedness of parity affect the same processing stage. In all analyses of the RT data, markednesscongruent and incongruent markedness stimulusresponse mapping for both even and odd numbers were grouped together according to markedness congruency: All markednesscongruent and markedness incongruent conditions could be separated by one straight line in all analyses. No other single factor such as magnitude, parity, or an interaction with response hand was effective in determining the structure in any analysis. Thus, the nonmetric MDS strongly supports the idea that congruency of the markedness of the parity of a given number and the markedness of response codes determine similarity of response patterns in this analysis. The SSA of the rightleft RT difference further confirms these conclusions. For all notations, even numbers (for which rightleft RT difference means compatibleincompatible condition) and odd numbers (for which rightleft RT difference means the reverse: incompatiblecompatible condition) are fully separated. Further, with respect to positive numbers, magnitude seemed to determine similarity structure, while for other notations no magnitude effect could be observed in the SSA structure. In the case of the degenerate solutions for negative Arabic numerals and number words, it can nevertheless be stated that magnitude did not play a role in the SSA structure for these notations within the current range of stimuli. In sum, it is remarkable, how much the results of the SSA differ from the results of the conventional ANOVA and regression analyses. Finally, the current SSA seems to confirm the idea of Brysbaert (1995) and others that zero is not part of the mental number line. For both negative and positive Arabic numerals in both response codes, right- and left-hand key, zero could be separated easily from all other numbers. For no other number could such an exceptional location could be consistently observed. On the use of multidimensional scaling techniques with RT data as compared to ordinary RT analyses. Average-based analyses (i.e., ANOVA and regression analyses) have been used mostly for RT or error data, while multidimensional scaling techniques have been applied to other off-line data (Miller & Gelman, 1983; Miller & Stigler, 1987, 1991; Shepard et al., 1975). Although Sternberg (1969) suggested analysing the correlational pattern of RT data, this has rarely since been done. The regression analysis of the RT and error data provided evidence for a SNARC effect in all notations (at least in the analyses of Lorch & Myers, 1990, and of the error data); but a MARC effect could only be observed for number words in all analyses and

854

NUERK, IVERSEN, WILLMES

(marginally) for positive Arabic numerals in the ANOVA analyses. In striking contrast, the MARC effect was obtained for all notations in the SSA, whileexcept for a possible tendency in the dRT MDS for positive Arabic numeralsno SNARC effect could be observed. This discrepancy can have two possible sources. First, the MDS with a regional interpretation of the spatial solution obtained is probably better suited to detect categorical effects such as the MARC effect than gradual effects like the SNARC effect.2 Second, the MARC effect may vary in a more systematic fashion between subjects than the SNARC effect. This may lead to a more stable pattern of correlation in the MDS while in the general linear model this interindividual variation contributes to error variance. First, the MARC effect is a categorical effect, which is supposed to be homogeneous for all members within one category (i.e., all within-category correlations should be higher than between-category correlations). Thus, the optimal correlational pattern is clear and may not be very sensitive to measurement errors. For the SNARC effect, which is a linear effect, the situation is different. For example, how should the answer to 5 (apart from the MARC effect) be correlated with 1 or 8 for the dRT data according to the SNARC effect? The numbers 1 and 5 should not be as highly (positively) correlated as 1 and 2, but still higher than 1 and 8. Thus, if one would expect medium size correlations in medium distance ranges between numbers, the MDS result for the SNARC effect may be more prone to measurement errors than the MDS results for the MARC effect. In summary, the MDS may be better suited to detect the reliability of the categorical MARC effect, while RT analysis may be better suited to examine the linear SNARC effect. Support for this argument comes from the average RT analyses. While a SNARC effect was obtained for all notations in the regression analysis with the Lorch and Myers (1990) method, the categorical interaction in the ANOVA between response hand and magnitude only reached significance for number words, and not for positive and negative Arabic numerals. The linear trend that is revealed in the regression analysis is not specifically tested in the categorical ANOVA, making the latter too conservative. The second argument may, conceptually, be equally important. Individuals may differ systematically with respect to their use of linguistic markedness. If such systematic differences between participants equally affect the markedness congruency for all numbers and both hand responses, a stable correlational pattern will be observed: for participants prone to the MARC effect, all MARC-incongruent trials may be slow as compared to MARC-congruent trials. For other participants this effect may play a lesser role. Thus, for the correlations over participants between, for example, two MARC incongruent trials, these systematic differences would result in quite high and stable correlations. Consequently, a stable MDS structure will be obtained. For an ANOVA (and the regression analysis), these individual differences in the MARC effect are conceptualized as error variance, because the ANOVA assumes that an independent factor should affect all replications in a cell of the experimental design (i.e., all
We wish to thank Wim Fias for many helpful suggestions and discussions concerning this article. In particular, he suggested performing a median split into a fast and slow half of participants. We also had a fruitful discussion about the issue of discrepancy between nMDS and ANOVA/regression analysis, in which he suggested the first argument with regard to that discrepancy. Finally, his comments were extremely helpful in strengthening the theoretical impact of the paper by relating the format-specificity of the MARC effect to current models of number processing and describing its functional locus more explicitly.
2

Downloaded by [University of St Andrews] at 04:22 24 April 2012

NOTATIONAL MODULATION OF SNARC AND MARC EFFECT TABLE 3 Overview of the SNARC and MARC effects obtained in this within-participant design for the different notations in the different analyses SNARC effect Positive Negative Number Arabic Arabic words () () () MARC effect Positive Negative Number Arabic Arabic words ()

855

Type of analysis RT ANOVA RT regression Error ANOVA Error regression SSA (abs. RT) SSA(dRT left - right)

Downloaded by [University of St Andrews] at 04:22 24 April 2012

Note: With regard to notation, the SNARC effect was most consistently found for positive Arabic numerals and number words; however, positive Arabic numerals was the only notation that tended to show a SSA-scaling pattern in accord with a SNARC-like ordering of numerals corresponding to magnitude. The MARC effect was most consistently obtained for number words. With regard to analyses methods, SNARC effects were obtained in the conventional ANOVA and regression analyses while MARC effects were reliably obtained for all notations in the SSA (see text for discussion).

participants) equally. Thus, the same individual differences that may lead to stable correlations in the MDS may increase error variance in the ANOVA and regression analysis and therefore lead to the different pattern of results in the two analyses.

GENERAL DISCUSSION
This study had three goals: First, we wished to examine the notational dependency of the SNARC and the hypothesized MARC effect in a within-participant design. We highlight the most important results of this study with regard to parity (MARC effect) and magnitude (SNARC effect; for an overview, see Table 3). Since the hypothesized MARC effect was found in all analyses for number words and still in some analyses for positive and negative numbers, we elaborate on a possible account for the MARC effect and its notation dependency. Additionally, we discuss how the results of this study constrain current models of number processing. With regard to the SNARC effect, we discuss particularly the results for negative numbers for which the SNARC effect tends to follow the absolute rather than the relative (negative) values. Finally, the role of the number zero is not clear in number processing research. Sometimes it is included in parity judgement tasks and sometimes it is excluded because of its special status. The nonmetric MDS results in this study provide convincing evidence that zero is indeed different from all other numbers in the parity judgement task.

On the notation dependency of the SNARC effect and the MARC effect
SNARC effect. For RT, significant SNARC effects were obtained for positive Arabic numerals and number words, but only a marginally significant SNARC effect for negative

856

NUERK, IVERSEN, WILLMES

Arabic numbers. In the error analyses, however, SNARC-like interactions (specifically tested in a linear-by-linear contrast) were observed for all three notations. For the first time, SNARC effects could be shown in the error analyses also for all notations (however, in a fixed-effects regression model, because many participants had zero errors). In the SSA, however, a tendency towards a SNARC scaling could only be observed for positive Arabic numbers in the RT difference SSA. For all other SSAs, there was no indication of a SNARC effect. So, in sum, the most reliable effects were found for positive Arabic digits: They were the only notation in which SNARC-like scaling effects could be observed. For number words, consistent SNARC effects were observed in the average RT and error ANOVAs and regression, but no indication of the SNARC effect could be obtained in the SSAs. The least reliable SNARC effects were obtained for negative numbers, which only reached conventional significance level in the error analyses, and not in any other analyses. However, although the SNARC effect for positive Arabic integers was more pronounced and more stable, the common RT SNARC effects for positive and negative Arabic integers did not differ significantly (as indexed by the regression weights). This was the case because the SNARC effect for negative numbers was marginally significant. The tendency of a SNARC effect, which was observed for negative Arabic numerals, had the same orientation with respect to their absolute value rather than their raw negative value. Had we computed the individual SNARC slopes with respect to the negative rather than the absolute value of negative numbers we would have obtained significant differences, t(22) = 3.00, p < .01. These patterns of results imply that the mental number line does not simply continue into negative Arabic numerals as in standard mathematical function graphs (at least for block-wise presentation as in this study). Two possible reasons for the less consistent SNARC effects for the negative Arabic numerals can be suggested. First, the minus sign, which is irrelevant for the parity decision, may nevertheless require additional processing capacity that masks the SNARC effect. The parity decision for negative numerals indeed tended to be slightly slower than that for positive Arabic integers (see Table 1). This hypothesis is, however, not consistent with the data of Fias, Lauwereyns, and Lammertyn (2001). They even obtained SNARC effects for irrelevant nonattended digits when stimuli other than these digits had to be processed in some adequate primary tasks. Given this SNARC effect for irrelevant digits , it seems unlikely that nonspecific interference from the irrelevant minus sign masks the SNARC effect when the digit themselves are still relevant. The second hypothesis is that negative Arabic numerals might evoke two spatialnumerical associations: one number line corresponding to the absolute values of the negative numerals, with 1 being on the far left and 8 being on the far right, which isaccording to our observed trendpredominant in most participants. A second less dominant number line might correspond to the spatialnumerical association of the raw negative numerals with 8 placed far left and 1 far right. The hypothesis that access to multiple number lines (or multiple access to one number line) is responsible for the less reliable results for negative Arabic numerals is consistent with recent data of Nuerk, Weger, and Willmes (2001, 2002) with regard to two-digit magnitude comparison. Nuerk and colleagues could show that the magnitude of the decade and the unit digits of two-digit numbers separately activate their respective magnitude representations. Thus, if multiple activation of different magnitude representations is possible, negative numbers may well activate both their absolute and their raw negative value to a certain extent.

Downloaded by [University of St Andrews] at 04:22 24 April 2012

NOTATIONAL MODULATION OF SNARC AND MARC EFFECT

857

Downloaded by [University of St Andrews] at 04:22 24 April 2012

MARC effect. MARC effects have been obtained for every notation in the SSA in this study. However, the notational differences in the RT and error analyses revealed a surprisingly different picture: The MARC effect was most reliably and strongly found for the verbal notation in all analyses. For positive Arabic numerals, MARC effects could only be significantly obtained in the error ANOVA and marginally in the RT ANOVA (besides the SSA). For negative Arabic integers, no MARC effect was observed in the average-based analyses. The nonmetric MDS results confirming the MARC effect in all analyses for all notations are surprising, as the partitioning of space according to linguistic markedness congruency was obtained even for positive and negative Arabic numerals for which no MARC effect was observed in the RT difference analyses. Following the logic of Sternbergs (1969) additivefactor method for correlational analysis, it is reasonable to assume that the markedness of the response codes and the markedness of parity influence the same processing stage because the association between the markedness of response codes and parity determines the scaling structure for all notations. We had not hypothesized that stimulusresponse markedness congruency would determine scaling structure for Arabic integers, and we were surprised by the clarity of this result (see in particular Figure 5 for the RT differences). In our opinion this pattern of results clearly indicates that the results obtained with one analysis method may not be conclusive for another analysis method. Finally, as one alternative explanation (see Footnote 2) for the MARC effect one could assume that it is just a peripheral instructionresponse compatibility effect: Press left if response is even is incompatible whatever the type of numbers. Indeed, this explanation could account for our pattern of results. However, this account is in our view inconsistent with the data reported by Hines (1990): In his samedifferent parity judgement task, participants had to decide whether or not two presented numbers had the same or different parities. The same responses should be made with the right hand and the different responses with the left hand. The critical difference with regard to the MARC effect is the difference between same responses to two odd versus two even numbers. The peripheral response (right hand) as well as the response decision same was identical in both conditions. Nevertheless, participants took much longer to respond same to two odd numbers than to two even numbers. In our opinion, a peripheral instructionresponse account of the MARC effect is incongruent with these data. However, a markedness congruency account would just predict that the markedness-congruent condition sameright (both unmarked) is slower than the markednessincongruent condition (differentright; markedunmarked). On the functional locus of the MARC effect with regard to current models of number processing (see Footnote 2).
Markedness may be a property more strongly associated with some internal linguistic representation(s) than with more conceptual, nonlinguistic representations such as digits. (Hines, 1990, p. 46).

No model of number processing has yet incorporated an explanation for the MARC effect obtained in this study. The only explanation available so far is that the congruency of linguistic markedness of the parity (oddeven) of a given number stimulus and of the linguistic markedness of response (leftright) facilitates performance as compared to markedness-incongruent

858

NUERK, IVERSEN, WILLMES

stimulusresponse mappings. In line with Hines (see above) we have argued that linguistic markedness as an abstract verbal concept may influence verbal notations more strongly (in a similar way that the positive Arabic notation seems to produce stronger SNARC effects). We elaborate on this assumption a little more below. The notational modulation of the MARC effects was quite clear in our study: The MARC was consistently much stronger and more reliable for verbal notation than for other notations. Any model trying to explain the MARC effect should also account for this notational modulation. The notation dependence of the MARC effect is hard to reconcile with the assumption that the underlying semantic representation (i.e., possibly linguistic markedness) exerts its influence on one central semantic parity representation equally for all notations. We now discuss briefly the notation specificity of the MARC effect with respect to three different models: the triple code model and its successors (Dehaene, 1992; Dehaene & Cohen, 1995, 1997), the modular model of McCloskey and colleagues (McCloskey, 1992; McCloskey, Macaruso, & Whetstone, 1992) and the encoding complex model of Campbell and Clark (Campbell, 1992, 1994; Campbell & Clark, 1988, 1992; Clark & Campbell, 1991). First, in the model of Dehaene, parity is retrieved from the visual Arabic number form. If this were the case, number words must be transformed to an Arabic number from which their parity, which is linked to markedness, could be retrieved. In this case, however, we should have observed identical parity (markedness) effects for Arabic numerals and for number words or even larger effects for Arabic numerals because Arabic integers are assumed to be closer linked to parity (cf. Dehaene et al., 1993). This was not the case: MARC effects were smaller for Arabic notation. In the modular model of McCloskey, parity retrieval is also supposed to rely on one single format: the abstract semantic representation (e.g., McCloskey et al., 1992). However, McCloskey and colleagues to our knowledge have not been explicit about the retrieval of parity. For example, Sokol, Goodman-Schulman, and McCloskey (1989, p. 108) point out that our model does not require that effects of oddeven status may be limited to calculation tasks. For example, effects of this variable could even be taken to suggest that the internal numerical representations posited by our model reflect oddeven status as well as numerical proximity (although we do not find this interpretation particularly appealing). However, the MARC effect as an interaction between numerical comprehension and output must necessarily be transformed into the abstract semantic number representation according to McCloskey. Even if parity is then retrieved via the number fact system or the calculation system, parity cannot be modulated by format, specifically because it has already been represented independent of notation. Format specificity could occur with main effects (words are slower than Arabic numerals), and with sophisticated interpretations even some interactions with notation may be explained (see, e.g., McCloskey et al., 1992). However, we are unable to see how format-specific interactions with the output such as the MARC effect are consistent with the models assumption of a central abstract semantic number representation that is obligatorily processed before any output is created. In short, the format specificity of the MARC effect is not well accounted for by the modular model. The only one among these three models that can in our opinion incorporate the MARC effect is the encoding complex theory of Campbell et al. The basic assumption of [their] approach is that number concepts and skills are based on modality and format-specific mental codes that are interconnected in a complex and highly integrated associative structure

Downloaded by [University of St Andrews] at 04:22 24 April 2012

NOTATIONAL MODULATION OF SNARC AND MARC EFFECT

859

(Campbell & Clark, 1992). Campbell and colleagues normally use the interaction effect of notation with another variable (e.g., type of errors) in calculation tasks to argue for their theory and as counter-arguments against the modular theory (Campbell, 1992, 1998, 1999; Campbell & Clark, 1992). In contrast to calculation tasks, the parity judgement task is much simpler to perform and easier to interpret. The notational modulation of the MARC effect isas are other interactions with notationconsistent with the assumption that format-specific codes and skills are used throughout a given task. In combination with our assumption that the MARC effect is modulated by a linguistic (verbal) attribute, the linguistic markedness, it is consistent with the encoding complex perspective that the MARC effect is more pronounced and more stable for verbal numeralsthat is, number words. Finally, we wish to elaborate a little more on how our account of the MARC effect may incorporate its notation dependency. Above, we have argued that linguistic markedness as a verballinguistic concept may activate verbal representations more strongly than nonverbal representations. What does verbal in this context mean? Is it the same verbal code that is used, for example, for stored multiplication facts in the model of Dehaene (see Cohen & Dehaene, 2000)? Furthermore, does it derive from a necessary processing stage or is it merely a side effect deriving from subvocal verbalization? The answers to these questions are speculative and call for further empirical investigation. First, with regard to the question of verbal codes, one needs a more fine-grain verbalprocessing account than most number-processing models possess. Most models simply assume that if numerals are presented in verbal form (number words), verbal-processing routes or modules must or can be used (e.g., Cipolotti & Butterworth, 1995; Fias, 2001, for recent discussions). However, the nature of this verbal processing is not very well specified. To achieve this, one needs to explore the possible source of the MARC effect in languageprocessing models. However, many language-processing models lack conceptual levels and have a lexical level as the highest implemented level. A model that incorporates a conceptual level is the language production model of Levelt and co-workers (e.g., Levelt, Roelofs, & Meyer, 1999). The concept of linguistic markedness is that most adjectives possesses an opposite form, with one of them being the basic form (the unmarked form) and the other being the derived form (the marked form). One would probably localize this selection at the highest level of the Levelt model, the conceptual preparation of lexical concepts (see Levelt et al., 1999, p. 3). If one tried to relate multiplication facts stored by verbal rote learning, one would probably locate this verbal process much lower in the Levelt model, possibly at the level of phonological encoding in the mental lexicon. Thus, although one might believe that the sources of the MARC effect and of the learned multiplication facts are both verbal they appear to be on very different levels of verbal processing. These considerations also allow more specific hypotheses about the derivation of the MARC effect. As linguistic markedness is a high-level concept, it is unlikely that it is strongly activated by a relatively low level process such as subvocal articulation. Rather, we believe that two lexical concepts are activated: those of even-ness (or odd-ness) and right-ness (or leftness) in this study as well as those of same-ness (or different-ness) and right-ness (or left-ness) in the Hines (1990) study. If two associated concepts share the same attribute (marked, unmarked) of linguistic markedness in the conceptual level of Levelt et al.s (1999) model, a response is facilitated while different markedness concepts may lead to interference. The markedness effect for number words may be stronger because number words activate the

Downloaded by [University of St Andrews] at 04:22 24 April 2012

860

NUERK, IVERSEN, WILLMES

Downloaded by [University of St Andrews] at 04:22 24 April 2012

verballexical entry associated with them rather strongly while Arabic numerals may do so to a lesser extent (note that the nMDS also revealed MARC effects for Arabic numerals). Arabic numerals may rather activate a representation (be it the visual number form as suggested by Dehaene or an abstract semanticmagnitude and parityrepresentation) that is not verbal lexical and helps the participant to decide whether the number is odd or even. MARC effects for Arabic notation may therefore rely on the coactivation of the verbal representation of this number. In contrast, number words also seem to activate their verbal lexical concepts that are associated with a certain markedness. If this linguistic markedness account is valid, markedness-related parity effects should therefore always be stronger for number words than for Arabic numbers. For instance, in the samedifferent-parity judgement tasks, even stronger odd effects should be obtained for number words than for digits. However, even for digits, the markedness effect can be tested. Attributes like symmetry/asymmetry have a clear markedness distinction. Although the markedness representation of parity may not totally disappear in a vertical symmetry judgement task about Arabic digits (as magnitude representation does not disappear in parity judgement task), the judgement symmetric may be faster for the (unmarked) right hand, and the judgement asymmetric may be faster for the (marked) left hand. In sum, the linguistic markedness account of the MARC effect and its notational modulation is currently the only account for the MARC effect that is also consistent with the results obtained by Hines (1990) in a samedifferent parity judgement task. Furthermore, the linguistic markedness account provides testable and falsifiable hypotheses for future research. On the special status of zero and five. Zero has sometimes been included in the analyses of parity judgement tasks and sometimes not (see introduction and Hines, 1990, vs. Dehaene et al., 1993), and the question whether or not to include zero is still controversial. Based on the magnitude effect obtained, Brysbaert (1995) took a clear standpoint on this controversy and claimed that zero is not part of the mental number line. In a parity judgement task, participants often are not sure about the parity status of zero and need to be introduced to the concept that it is an even number (which is true in terms of mathematical group theory). We investigated the role of zero with multidimensional scaling techniques for the first time in a parity judgement task. Our SSA results clearly showed that zero is separated from any other positive and negative Arabic numeral for both right- and left-hand responses while the other odd and even numbers (in association with the congruent or incongruent response hand) were always grouped together. This result is inconsistent with the assumption that the parity status of zero is that of other even numbers. Therefore, we suggest that zero should no longer be included in parity judgement experiments. At least, it should be controlled whether the effects and null effects obtained in a given parity judgement task do still prevail if zero is excluded. SNARC, MARC, or odd effects whose significance only rests on the inclusion of zero canwhen our SSA results are consideredhardly be generalized to magnitude or parity representation in general. In physical and numerical magnitude comparison studies, it has been claimed that the number 5 has a special status (Tzelgov, Meyer, & Henik, 1992). Tzelgov et al. suggested that the digits 14 or 24 are automatically considered to be small, while the digits 69 or 68 are automatically considered to be large. Indeed, they could show that in a physical judgement task, the size congruity effect differed for trials with and without five, perhaps because of the

NOTATIONAL MODULATION OF SNARC AND MARC EFFECT

861

special status of five as the neutral middle in the one-digit number range of the mental number line. In our study, we have found no evidence that five is represented differently than other odd numbers. The RT and error SNARC effects revealed no special status for number 5 and in the nonmetric MDS; it was always grouped with the other MARC-congruent or MARCincongruent odd numbers. Thus, the special status that the number 5 may well have in magnitude comparison tasks was not observed for parity judgement in any notation in this study.

CONCLUSIONS
Downloaded by [University of St Andrews] at 04:22 24 April 2012
First, this study has established that parity is associated with response code (evenright, odd left). This MARC effect is notation dependent: It is most reliably observed for verbal notation and less for Arabic nonlinguistic notation. Our account of this pattern of results is the linguistic markedness hypothesis: If the parity of a number has a specific markedness (even unmarked, oddmarked), responses tend to be faster for some notations when the response code has the same markedness (rightunmarked, leftmarked). The notation specificity of the MARC effect in the different analyses is hard to reconcile with models that assume a central abstract representation of parity for all notations alike. Second, multidimensional scaling techniques can be helpful in exploring the similarity structure of numbers. Markedness association determined the correlational grouping structure also for positive and negative Arabic integers, while no respective RT differences were found. For magnitude, the opposite was true: The SNARC effect was obtained in the analysis of mean RT differences and in the error analyses. However, in the nonmetric MDS, magnitude only influenced the pattern of similarities for positive Arabic integers. Finally, multidimensional scaling indicated that the zero is represented differently from all other numbers when participants are asked about parity. Therefore, it is recommended that effects obtained in parity judgement tasks (e.g., SNARC, MARC, odd effect) should at least be also tested without being checked for the impact of zero. Effects that rely on the inclusion of zero canin our opinionnot easily be generalized to number representation in general.

REFERENCES
Armstrong, S. L., Gleitman, L. R., & Gleitman, H. (1983). What some concepts might not be. Cognition, 13, 263308. Ashcraft, M. H. (1992). Cognitive arithmetic: A review of data and theory. Cognition, 44, 75106. Bchthold, D., Baumller, M., & Brugger, P. (1998). Stimulus-response compatibility in representational space. Neuropsychologia, 36, 731735. Berch, D. B., Foley, E. J., Hill, R. J., & McDonough Ryan, P. (1999). Extracting parity and magnitude from Arabic numerals: Developmental changes in number processing and mental representation. Journal of Experimental Child Psychology, 74, 286308. Beringer, J. (1993). Handbook for ERTS: Experimental Run Time System, Version 3.00. Frankfurt am Main, Germany: BeriSoft Cooperation. Borg, I., & Lingoes, J. (1987). Multidimensional similarity structure analysis. New York: Springer. Brysbaert, M. (1995). Arabic number reading: On the nature of the numerical scale and the origin of phonological recoding. Journal of Experimental Psychology: General, 124, 434452. Buckley, P. B., & Gillman, C. B. (1974). Comparisons of digits and dot patterns. Journal of Experimental Psychology, 103, 11311136. Butterworth, B. (1999). The mathematical brain. London: Macmillan.

862

NUERK, IVERSEN, WILLMES

Campbell, J. I. D. (1992). In defense of the encoding-complex approach: Reply to McCloskey, Macaruso, & Whetstone. In J. I. D. Campbell (Ed.), The nature and origins of mathematical skills (pp. 539556). Amsterdam: Elsevier. Campbell, J. I. D. (1994). Architectures for numerical cognition. Cognition, 53, 144. Campbell, J. I. D. (1998). Linguistic influences in cognitive arithmetic: Comment on Noel, Fias, and Brysbaert (1997). Cognition, 67, 353364. Campbell, J. I. D. (1999). The surface form problem size interaction in cognitive arithmetic: Evidence against an encoding locus. Cognition, 70, B25B33 Campbell, J. I. D., & Clark, J. M. (1988). An encoding-complex view of cognitive number processing: Comment on McCloskey, Sokol, and Goodman (1986). Journal of Experimental Psychology: General, 117, 204214. Campbell, J. I. D., & Clark, J. M. (1992). Cognitive number processing: An encoding complex perspective. In J. I. D. Campbell (Ed.), The nature and origins of mathematical skills (pp. 457491). Amsterdam: Elsevier. Cipolotti, L., & Butterworth, B. (1995). Toward a multiroute model of number processing: Impaired number transcoding with preserved calculation skills. Journal of Experimental Psychology: General, 124, 375390. Clark, J. M., & Campbell, J. I. D. (1991). Integrated versus modular theories of number skills and acalculia. Brain and Cognition, 17, 204239. Cohen, L., & Dehaene, S. (2000). Calculating without reading: Unsuspected residual abilities in pure alexia. Cognitive Neuropsychology, 17, 563583. Dehaene, S. (1989). The psychophysics of numerical comparison: A reexamination of apparently incompatible data. Perception & Psychophysics, 45, 55766 Dehaene, S. (1992). The varieties of numerical abilities. Cognition, 44, 142. Dehaene, S., & Akhavein, R. (1995). Attention, automaticity, and levels of representation in number processing. Journal of Experimental Psychology: Learning, Memory, and Cognition, 2, 314326. Dehaene, S., Bossini, S., & Giraux, P. (1993). The mental representation of parity and number magnitude. Journal of Experimental Psychology: General, 122, 371396. Dehaene, S., & Cohen, L. (1995). Towards an anatomical and functional model of number processing. Mathematical Cognition, 1, 83120. Dehaene, S., & Cohen, L. (1997). Cerebral pathways for calculation: Double dissociation between rote verbal and quantitative knowledge of arithmetic. Cortex, 33, 219250. Dehaene, S., Dupoux, E., & Mehler, J. (1990). Is numerical comparison digital? Analogical and symbolic effects in twodigit number comparison. Journal of Experimental Psychology: Human Perception and Performance, 16, 626 641. Fias, W. (2001). Two routes for the processing of verbal numbers: Evidence from the SNARC effect. Psychological Research, 12, 415423. Fias, W., Brysbaert, M., Geypens, F., & dYdewalle, G. (1996). The importance of magnitude information in numerical processing: Evidence from the SNARC-effect. Mathematical Cognition, 2, 95110. Fias, W., Lauwereyns, J., & Lammertyn, J. (2001). Irrelevant digits affect feature-based attention depending on the overlap of neural circuits. Cognitive Brain Research, 12, 415423. Foltz, G. S., Poltrock, S. E., & Potts, G. R. (1984). Mental comparison of size and magnitude: Size congruity effects. Journal of Experimental Psychology: Human Perception and Performance, 10, 442453. Guttman, L. (1968). A general nonmetric technique for finding the smallest coordinate space for a configuration of points. Psychometrika, 33, 469506. Hines, T. M. (1990). An odd effect: Lengthened reaction times for judgments about odd digits. Memory and Cognition, 18, 4046. Hinrichs, J. V., Yurko, D. S., & Hu, J.-M. (1981). Two-digit number comparison: Use of place information. Journal of Experimental Psychology: Human Perception and Performance, 7, 890901. Levelt, W. J. M., Roelofs, A., & Meyer, A. S. (1999). A theory of lexical access in speech production. Behavioral and Brain Sciences, 22, 175. Lingoes, J. C. (1973). The Guttman-Lingoes nonmetric program series. Ann Arbor, MI: Mathesis Press. Lorch, R. F., Jr., & Myers, J. L. (1990). Regression analyses of repeated measures data in cognition research. Journal of Experimental Psychology: Learning, Memory, and Cognition, 16, 149157. Lyons, J. (1969). Semantics. Cambridge: Cambridge University Press. McCloskey, M. (1992). Cognitive mechanisms in numerical processing: evidence from acquired dyscalculia. Cognition, 44, 10757.

Downloaded by [University of St Andrews] at 04:22 24 April 2012

NOTATIONAL MODULATION OF SNARC AND MARC EFFECT

863

Downloaded by [University of St Andrews] at 04:22 24 April 2012

McCloskey, M., Macaruso, P., & Whetstone, T. (1992). The functional architecture of numerical processing mechanisms: Defending the modular model. In J. I. D. Campbell (Ed.), The nature and origins of mathematical skills (pp. 493537). Amsterdam: Elsevier. Miller, K. F., & Gelman, R. (1983). The childs representation of number: A multidimensional scaling analysis. Child Development, 54, 14701479. Miller, K. F., & Stigler, J. W. (1987). Counting in Chinese: Cultural variation in a basic cognitive skill. Cognitive Development, 2, 279305. Miller, K. F., & Stigler, J. W. (1991). Meanings of skill: Effects of abacus expertise on number representation. Cognition and Instruction, 8, 2967. Moyer, R. S., & Landauer, T. K. (1967). Time required for judgments of numerical inequality. Nature, 215, 1519 1520. Nuerk, H. C., Weger, U., & Willmes K. (2002) Decade breaks in the mental number line? Putting the tens and units back in different bins. Cognition, 82 , B2533. Nuerk, H. C., Weger, U., & Willmes, K. (in press). A unit-decade compatibility effect in German number word. Current Psychology Letters: Behavior, Brain, & Cognition, 7, 1938. Restle, F. (1970). Speed of adding and comparing numbers. Journal of Experimental Psychology, 83, 3245. Reynvoet, B., & Brysbaert, M. (1999). Single-digit and two-digit Arabic numerals address the same semantic number line. Cognition, 72, 191201. Shepard, R. N., Kilpatric, D. W., & Cunningham, P. (1975). The internal representation of numbers. Cognitive Psychology, 7, 82138. Sherman, M. A. (1973). Bound to be easier? The negative prefix and sentence comprehension. Journal of Verbal Learning and Verbal Behavior, 12, 7684. Sherman, M. A. (1976). Adjectival negation and the comprehension of multiply negated sentences. Journal of Verbal Learning and Verbal Behavior, 15, 143157. Sokol, S. M., Goodman-Schulman, R., & McCloskey, M. (1989). In defense of a modular architecture for the number-processing system: Reply to Campbell and Clark. Journal of Experimental Psychology: General, 118, 105 110. Sternberg, S. (1969). The discovery of processing stages: Extension of Donders method. Acta Psychologica, 30, 276 315. Sternberg, S. (1998). Discovering mental processing stages: The method of additive factors. In D. Scarborough & S. Sternberg (Eds.), An invitation to cognitive science: Methods, models, and conceptual issues (2nd ed., pp. 703863). Cambridge, MA: MIT Press. Tzelgov, J., Meyer, J., & Henik, A. (1992). Automatic and intentional processing of numerical information. Journal of Experimental Psychology: Learning, Memory, and Cognition, 18, 166179. Willmes, K., & Iversen, W. (1995 April). On the Internal Representation of Number Parity. Paper presented at the Spring Annual Meeting of the British Neuropsychological Society, London. Zimmer, K. (1964). Affixed negation in English and other languages: An investigation of restricted productivity. Word, 20, 2, Monograph No. 5. Original manuscript received 15 May 2001 Accepted revision received 24 April 2003 PrEview proof published online 3 September 2003

You might also like