Cognition 118 (2011) 97–113 Contents lists available at ScienceDirect Cognition journal homepage: www.elsevier.com/locate/COGNIT The size congruity effect: Is bigger always more? Seppe Santens ⇑, Tom Verguts Ghent University, Department of Experimental Psychology, H. Dunantlaan 2, B-9000 Ghent, Belgium a r t i c l e i n f o Article history: Received 1 October 2009 Revised 19 October 2010 Accepted 20 October 2010 Keywords: Size congruity Numerical Stroop Numerical cognition Number processing a b s t r a c t When comparing digits of different physical sizes, numerical and physical size interact. For example, in a numerical comparison task, people are faster to compare two digits when their numerical size (the relevant dimension) and physical size (the irrelevant dimension) are congruent than when they are incongruent. Two main accounts have been put forward to explain this size congruity effect. According to the shared representation account, both numerical and physical size are mapped onto a shared analog magnitude representation. In contrast, the shared decisions account assumes that numerical size and physical size are initially processed separately, but interact at the decision level. We implement the shared decisions account in a computational model with a dual route framework and show that this model can simulate the modulation of the size congruity effect by numerical and physical distance. Using other tasks than comparison, we show that the model can simulate novel findings that cannot be explained by the shared representation account. ! 2010 Elsevier B.V. All rights reserved. 1. Introduction In order to act adaptively in a constantly changing environment, animals need to keep track of different quantities, including time, length, size, area, and volume. In recent evolutionary history, also abstract quantities (numbers) have been added to this list. Research in different disciplines has recently focused on understanding the cognitive foundations of these different quantities. A still unresolved issue, however, concerns the relation between them (for reviews, see Cantlon, Platt, and Brannon (2009), Cohen Kadosh, Lammertyn, and Izard (2008)). An effect that plays a pivotal role in this discussion is the size congruity effect (Besner & Coltheart, 1979; Henik & Tzelgov, 1982). In the original size congruity paradigm (Besner & Coltheart, 1979), two numbers are presented in different physical sizes, and the participant’s task is to choose the numerically larger (or smaller) stimulus. The size congruity effect entails that it is easier to pick the numerically larger (or smaller) of two digits when this is ⇑ Corresponding author. E-mail address: [email protected] (S. Santens). 0010-0277/$ - see front matter ! 2010 Elsevier B.V. All rights reserved. doi:10.1016/j.cognition.2010.10.014 also the physically larger (smaller) stimulus. Similarly, when the task is to pick the physically larger (or smaller) of the two digits, the task is easier if the physically larger stimulus is also numerically larger (Henik & Tzelgov, 1982). Size congruity effects are also found when only one stimulus is presented per trial. In this paradigm, the task is to judge whether the presented digit is larger or smaller than a fixed standard (e.g., standard 5) (Schwarz & Heinze, 1998; Schwarz & Ischebeck, 2003). The current paper addresses the origins of this effect using both computational and behavioral studies. The size congruity effect shows that numerical and physical (size) dimensions are not processed in complete independence. An interaction between different quantities must be assumed somewhere along the processing stream. This interaction can take place at the input, representation, decision, or output level (Verguts & Fias, 2008). It is still debated at which of these levels the size congruity effect originates. In the literature, two opposing accounts have been put forward (e.g., Cohen Kadosh et al., 2007; Schwarz & Heinze, 1998). Schwarz and Heinze (1998) have clearly identified both positions, so we take their definitions as examples. In the first account, the size congruity effect is taken as evidence for representational overlap between 98 S. Santens, T. Verguts / Cognition 118 (2011) 97–113 the dimensions of number and physical size. According to this account, it is the case that ‘‘[. . .] both, the digit’s physical size and its numerical value are first mapped onto an integrated internal analog representation, which is then processed further to activate the appropriate response’’ (Schwarz & Heinze, 1998, p. 1168). We will call this the shared representation account (Fig. 1A). In this account, size congruity originates at the representational level, i.e. a level where numerical and physical size would be jointly represented in an analog format. A stimulus configuration in which numerical size correlates positively with physical size is congruent (e.g., in the stimulus configuration 2 8, the numerical and physical size are both small for the left stimulus and both large for the right stimulus). A stimulus configuration in which the two dimensions correlate negatively, is incongruent (e.g., in the stimulus configuration 8 2). An instance of this shared representation theory is Walsh’s influential theory of magnitude (ATOM), according to which there are common metrics for different quantitative dimensions, including space, time, and number (Bueti & Walsh, 2009; Walsh, 2003). Many recent studies on congruity effects between numerical size and other quantities have been interpreted in terms of ATOM (e.g., Chiou, Chang, Tzeng, & Wu, 2009; Cohen Kadosh, Cohen Kadosh, & Henik, 2008; Cohen Kadosh & Henik, 2006; Cohen Kadosh et al., 2007; Kaufmann et al., 2005; Pinel, Piazza, Le Bihan, & Dehaene, 2004; Xuan, Zhang, He, & Chen, 2007). A theoretical problem with this proposal, and with the shared representation account more generally, is that the common metric for the shared dimensions has remained unspecified. Put simply, it is unclear whether 1 s (quantity in the time dimension) corresponds to 1 cm or 1 m (quantity in the length dimension); or to 1 g or 1 kg (quantity in the weight dimension); and so on. One exception is the proposal by Meck and Church (1983) for counting and time representations in rats. In their mode-control model, both counting and timing rely on a pacemaker that sends pulses to an accumulator. Depending on the task, the accumulator fills up to represent either time or the number of discrete elements (i.e., counting). They even proposed that a count corresponds to a passage of time of 200 ms (Meck & Church, Experiment 4). However, it is unclear how general this count-to-time mapping is. Furthermore, the model cannot easily be extended to other quantitative dimensions. The second account for explaining the size congruity effect suggests that it results from the fact that, although numerical and physical size are initially processed separately, the two separate processing pathways interact at the decision level. According to this second account it is the case that ‘‘[. . .] size and numerical information are first processed in parallel, functionally independent channels and that both of them can activate a specific ‘‘subresponse’’ ’’ (Schwarz & Heinze, 1998, p. 1168). We can think of these subresponses as the decision alternatives that are imposed by the task and we therefore call this account the shared decisions account (Fig. 1B). In a standard size congruity paradigm, the decision alternatives (e.g. left larger or right larger) have a one-to-one mapping to the responses (e.g. left hand or right hand). Therefore, dissociating between the decision level and the output level is neither possible nor relevant here. Importantly, the shared decisions account assumes that the size congruity effect is due to the fact that the two dimensions activate the same (in the congruent case) or different (in the incongruent case) task-relevant, discrete codes. Hence, according to this account, congruity can be defined only relative to the task at hand. For example, in a magnitude comparison task (but not necessarily in other tasks), the stimulus configuration 2 8 is congruent because the numerical and physical size both activate the right larger code. In contrast, the stimulus configuration 8 2 is incongruent because the numerical dimension activates the left larger code, whereas the physical dimension activates the right larger code. In a sense, the activation of the decision units also establishes a representation, namely of the decision that is made. This could cause confusion about what is exactly meant by a shared representation account. However, the definitions of Schwarz and Heinze (1998) are clear: In the shared representation account, the size congruity effect originates at the level where numerical and physical magnitude are jointly coded in an analog format. This information about Fig. 1. Schematic illustration of the shared representation account (A) and the shared decisions account (B). The origin of the size congruity effect according to each account is marked. S. Santens, T. Verguts / Cognition 118 (2011) 97–113 numerical and physical size is lost at the decision level. We reserve the term ‘‘shared representation’’ for a representation that codes numerical and physical size in a shared analog format, before task-specific operations are applied to it. Although many studies have discussed the interaction between numerical and physical size (e.g. Algom, Dekel, & Pansky, 1996; Ansari, Fugelsang, Dhital, & Venkatraman, 2006; Henik & Tzelgov, 1982; Kaufmann et al., 2005; Pinel et al., 2004; Schwarz & Ischebeck, 2003), few have been designed to directly oppose the shared representation and the shared decisions account. Schwarz and Heinze (1998) recorded event-related potentials (ERPs) from six electrodes to determine the locus of the size congruity effect. In their results, the first reliable effects of congruity at the frontal electrodes occur at 280 ms after stimulus presentation for the physical comparison task and at 368 ms for the numerical comparison task. The authors interpreted this finding as evidence for early interactions and hence for a shared representation. They also examined the lateralized readiness potential (LRP), an ERP component that is thought to reflect motor preparation and execution (Coles, 1989; Hackley & Miller, 1995; Masaki, Wild-Wall, Sangals, & Sommer, 2004). An initial deflection of the LRP towards the incorrect response in incongruent trials compared to congruent trials is considered as evidence for response competition between the correct response activated by the relevant dimension and the incorrect response automatically activated by the irrelevant dimension (e.g., Gevers, Ratinckx, De Baene, & Fias, 2006). Schwarz and Heinze (1998) interpreted the absence of such a deflection in their data as additional support for a shared representation. However, as also argued by Szücs and Soltész (2008), the absence of an initial LRP deflection cannot be taken as evidence in favor of a shared representation. If the irrelevant dimension is processed more slowly than the relevant dimension (e.g. because the relevant dimension is emphasized by task instructions), then it is possible that the irrelevant dimension does not cause an initial LRP deflection and yet slows down selecting the correct response. Szücs and Soltész (2008) also recorded ERPs in a size congruity paradigm. They found both early (between 150 and 250 ms after stimulus presentation) and late (between 300 and 430 ms) facilitation and interference effects of numerical and physical dimensions. They suggested that the size congruity effect appears at multiple levels of stimulus processing and response preparation. Cohen Kadosh et al. (2007) combined ERP measures with functional magnetic resonance imaging (fMRI) in both a numerical and a physical comparison task. A region of interest analysis of the fMRI data from the primary motor cortex showed an interference effect in the hemisphere ipsilateral to the hand for the correct response. It was concluded that congruity is not completely resolved until response initiation. In their ERP analyses, the authors found an initial deflection of the LRP when numerical distance was large, but not when numerical distance was small. It was argued that response competition plays a role only when cognitive load is low. Under a high cognitive load, a shared representation would cause the congruity effect. However, as discussed above, the absence of an LRP effect does not justify this 99 conclusion. A more general problem with the available ERP evidence is that there is no consensus on what constitutes an ‘‘early’’ or ‘‘late’’ ERP component. For example, both Schwarz and Heinze (1998) and Cohen Kadosh et al. (2007) categorize the congruity effect around 300 ms after stimulus onset as early, whereas Szücs and Soltész (2008) categorize this as a late component. In sum, the available behavioral and neuroimaging results are not conclusive in dissociating between the shared representation and shared decisions accounts. A crucial point that blurs the discussion is the fact that both accounts so far have remained underspecified. Only when both are specified in sufficient detail will it be possible to resolve this debate. As mentioned above, specifying the shared representation account has an inherent indeterminacy problem (what is the common metric for quantities?). On the other hand, the shared decisions account can be straightforwardly modeled with a dual route architecture. This approach has proven successful in congruity paradigms similar to numerical and physical size congruity. Cohen, Dunbar, and McClelland (1990) for example simulated the well-known Stroop effect (Stroop, 1935) with a computational model implementing a dual route architecture. In their model, an ‘‘ink color’’ and a ‘‘word’’ route both activate one of the two possible responses (‘‘red’’ or ‘‘green’’). An attentional bias strengthens the task-relevant route (i.e. ‘‘ink color’’ in a classic Stroop task) to ensure that the correct response is selected (e.g. response ‘‘red’’ when the word the ink color is ‘‘red’’). The automatically activated task-irrelevant route (‘‘word’’) causes facilitation in congruent trials (e.g. the word ‘‘red’’ printed in red) by activating the correct response unit and it causes interference in incongruent trials (e.g. the word ‘‘green’’ printed in red) by activating the incorrect response unit. Later, also other congruity effects were successfully modeled with a dual route architecture, including the Eriksen flanker effect (Cohen, Servan-Schreiber, & McClelland, 1992; Yeung, Botvinick, & Cohen, 2004), the Simon effect (Zorzi & Umiltà, 1995) and the SNARC effect (Gevers, Verguts, Reynvoet, Caessens, & Fias, 2006). In the current study, we took a similar approach to model the size congruity effect. For this purpose, we started from our earlier computational model of number processing (Verguts, Fias, & Stevens, 2005). This model simulates behavioral performance in a number of tasks, but here we focus on number comparison. The model uses place coding (i.e. number-selective coding) to represent numerical magnitude on two input layers, one for each of the to be compared numbers. These input layers then activate the output layer, which codes for the decision that has to be made. If the instruction is to choose the larger number, the units of the output layer code for ‘‘left digit is larger’’ and ‘‘right digit is larger’’ (if the responses are arranged left and right). After training the model on the comparison task, the pattern of weights between the input and output layers produces a comparison distance effect (for details, see Verguts et al., 2005). Here, we implement only the comparison mechanism of the Verguts, Fias, and Stevens model in a simplified way. To simulate the size congruity paradigm, we add to this a similar mechanism for physical size comparison. Like the model of Cohen 100 S. Santens, T. Verguts / Cognition 118 (2011) 97–113 et al. (1990), the task-relevant route (e.g. numerical size comparison) and the task-irrelevant route (e.g. physical size comparison) interact only at the decision level. At this level, the units that code for the decision alternatives (e.g. ‘‘left digit is larger’’ and ‘‘right digit is larger’’) are shared between the numerical and the physical dimension. Hence, the resulting model is an instance of the shared decisions account. Notably, the model is always trained on numerical and physical size comparison separately. Only in the test phase, the model is confronted with the size congruity paradigm. The distinction between the relevant dimension and the irrelevant dimension is made by simply attenuating the activation from the irrelevant dimension (for details, see Appendix B). With four behavioral experiments and four simulations, we show that the principle of physical and numerical routes operating independently but converging at the decision level can account for many aspects of the size congruity effect. In Experiment 1, we first replicate the typically observed interactions of numerical and physical distance with size congruity and test whether the shared decisions model can simulate these effects. In Experiments 2–4, we directly oppose the two accounts. In Experiment 2, it is tested whether a size congruity effect is present in both a magnitude judgment task and a parity judgment task because only the shared representation account predicts that it will be observed in both tasks. We further investigate the influence of the task on the size congruity effect in Experiment 3. Here, we use a close/far task (Santens & Gevers, 2008) to test whether modulations of the size congruity effect follow the predictions of the shared representation account or of the shared decisions account. In the first three experiments, numerical size is the relevant dimension and physical size the irrelevant dimension. In Experiment 4, we investigate whether the conclusions based on Experiment 3 still hold when changing the relevant and the irrelevant dimension. We use the same close/far task as in Experiment 3, but now with physical size as the relevant dimension and numerical size as the irrelevant dimension. 2. Experiment 1 The size congruity effect can be affected by manipulating the discriminability between the values on the relevant and irrelevant dimension (Pansky & Algom, 1999). Therefore, we calibrated the physical sizes of the stimuli used in this study in order to guarantee equal discriminability on the two dimensions. In Experiment 1, this was done separately for each participant on the first day of the experiment (see Appendix A). On the second day, we used the calibrated stimuli in a standard size congruity paradigm. 2.1. Method 2.1.1. Participants Eighteen students from Ghent University were paid to participate in an experiment of two times 1 h (for the calibration phase and the experimental phase, respectively), spread over two days within the same week. 14 participants were female, all but one were right-handed. Their mean age was 21 years (range: 18–24 years). All participants had normal or corrected to normal vision. They signed an informed consent prior to the experiment. 2.1.2. Stimuli and apparatus The experimental procedure was implemented using the Tscope library for the C programming language (Stevens, Lammertyn, Verbruggen, & Vandierendonck, 2006). Button presses were registered using a response box (RB730 with modified buttons, Cedrus Corp., San Pedro, CA). Stimulus pairs were presented on a 17 in. screen, white on a black background. The viewing distance was approximately 50 cm. Stimuli consisted of the Arabic numerals 1, 2, 7 and 8, printed in the Courier New font. These digits appeared in 4 different heights (a < b < g < h), determined in the calibration phase (see Appendix A). The mean heights in visual degrees for a, b, g and h were about 2.7", 5.4", 9.4", and 13.3", respectively. On each trial, two stimuli were presented, centered around the vertical midpoint and 7.4 visual degrees left and right from the fixation point, respectively. 2.1.3. Experimental procedure Participants had to perform a number comparison task on digits that varied in numerical and physical size. They were asked to ignore the physical size of the digits. The specific task instruction (‘‘choose the smaller’’ or ‘‘choose the larger’’) was the same as in the calibration phase and thus counterbalanced between subjects. A trial started with the presentation of a small square in the center of the screen that served as the fixation mark. There was a variable time interval before the stimulus appeared, with a duration randomly drawn from a uniform distribution with a minimum of 500 ms and a maximum of 1000 ms. After this interval, the stimuli appeared, one on the right and one on the left. There was a response deadline of 1500 ms. When the response was given or when 1500 ms had elapsed, there was a blank screen for 500 ms until the next trial started. All combinations of numerical (1, 2, 7, 8) and physical sizes (a, b, g, h) were presented, except for the combinations in which the numerical or the physical size was equal for the left and right stimulus. In this way, 12 ! 12 = 144 different stimulus pairs could be constructed. Each pair was presented once in each of the eight blocks in the experiment. This gave a total of 1152 trials. Between blocks, participants could take a short pause. 2.2. Results and discussion All correct trials entered the RT analyses. The results of these analyses are presented in Fig. 2. Mean RTs were calculated for each condition, collapsing over response instructions (choose the numerically larger or smaller). An ANOVA was run with a 2 (congruity: congruent or incongruent) ! 2 (distance on the (relevant) numerical dimension: small or large) ! 2 (distance on the (irrelevant) physical dimension: small or large) design with all factors treated as within subjects variables. To obtain an equal number of observations at each level of numerical distance, the two smallest distances (distances 1 and 5) were treated as small and the two largest distances (distances 6 S. Santens, T. Verguts / Cognition 118 (2011) 97–113 101 Fig. 2. Experiment 1: observed (panels A and C) and simulated (panels B and D) RT. Panels A and B: congruity plotted against distance on the relevant dimension. Panels C and D: congruity plotted against distance on the irrelevant dimension (Error bars represent 95% confidence intervals based on the guidelines by Loftus and Masson (1994). Error bars for the model data were too small to depict. To aid visual inspection of the effect sizes, the model data of all simulations are presented on the same scale and so are the behavioral data of all experiments. The range of values on the Y axis is the same for all simulations, whereas it changes between experiments for the behavioral data. Note that the data points in these graphs are marginal means. Therefore the triple interaction cannot be evaluated from this figure). and 7) were treated as large. A similar categorization was used for physical distance: Small distances were |a–b|, |b–g| and |g–h|; large distances were |a–g|, |a–h|, and |b– h|. There was a main effect of congruity [F(1, 17) = 45.42; p < .001; g2p ¼ :73]: Participants were faster to pick the numerically larger or smaller of two stimuli when their numerical and physical size were congruent than when they were incongruent. The main effect of numerical distance was also significant [F(1, 17) = 188.27; p < .001; g2p ¼ :92]: When the numerical distance was large, responses were faster than when the numerical distance was small. Importantly, there was a significant interaction between congruity and numerical distance [F(1, 17) = 15.64; p < .01; g2p ¼ :48]: The congruity effect (i.e. the absolute difference between RT for congruent pairs and RT for incongruent pairs) was larger for a small numerical distance than for a large numerical distance (see Fig. 2A). There was also a main effect of physical distance [F(1, 17) = 26.93; p < .001; g2p ¼ :61]: Overall, participants were faster to respond when physical distance was small than when it was large. The interaction between congruity and physical distance was also significant [F(1, 17) = 6.50; p < .05; g2p ¼ :28] and in the opposite direction as for relevant distance (see Fig. 2C). The interaction between numerical distance and physical distance and the three- way interaction were not significant (both F(1, 17) < 1; g2p < :05). On average, participants did not respond in time or made an error on only 1.8% of all trials. The same ANOVA used for the RT analysis was used for the analysis of the proportion of errors after arcsine transformation. The pattern of the RT analysis was reflected in the error rates analysis: in all cases, more errors were made in the slower conditions. There were significant main effects of congruity [F(1, 17) = 23.72; p < .001; g2p ¼ :58], numerical distance [F(1, 17) = 67.51; p < .001; g2p ¼ :80], and physical distance [F(1, 17) = 48.45; p < .001; g2p ¼ :74]. There were significant two-way interactions between congruity and numerical distance [F(1, 17) = 18.16; p < .01; g2p ¼ :52] and between congruity and physical distance [F(1, 17) = 5.40; p < .05; g2p ¼ :24]. In contrast to the RT analysis, there was also a two-way interaction between numerical and physical distance [F(1, 17) = 19.64; p < .001; g2p ¼ :54]: the numerical distance effect was larger when physical distance was large than when it was small. To sum up, we replicated the main findings reported by Schwarz and Ischebeck (2003) for the numerical comparison task. In particular, the size congruity effect is larger for small relevant distances than for large relevant distances; but it is larger for large irrelevant (physical dimension) distances 102 S. Santens, T. Verguts / Cognition 118 (2011) 97–113 than for small irrelevant distances. An interpretation of this difference will be provided in the next section. 3. Simulation 1 In Simulation 1, we show that the shared decisions account can be implemented in a computational model with a dual route architecture. We evaluate whether the model can simulate the effects of Experiment 1 and explain how these effects originate from the model. 3.1. Method As outlined in the introduction, our implementation of the shared decisions account has a numerical and a physical processing route that interact only at the decision level. These routes both implement the comparison mechanism of the Verguts et al. (2005) model. The model was trained separately for numerical size comparison and physical size comparison. During this training, the weights between the representation layers and the right larger and left larger decision units were adjusted in order to improve performance of the model on the numerical and the physical comparison tasks. After training, the model was tested using the 144 stimulus pairs from Experiment 1 that varied in numerical and physical size. To show that the dual route architecture of the model is sufficient to simulate the behavioral effects, we used a basic steady state model and did not implement the specificities of a response selection mechanism (for a similar approach, see Van Opstal, Gevers, De Moor, & Verguts, 2008). The amount of activation of a decision unit was interpreted as the amount of evidence in favor of the corresponding decision. We formalized the intuitive notion that the more evidence there is for one decision relative to the other one, the faster the response time (RT) will be (e.g., Ratcliff, 1978; Usher & McClelland, 2001) by defining simulated RT as an inverse function of the absolute difference in activation of the two decision units. See Fig. 3 for a graphical illustration of the model and Appendix B for details of the implementation. 3.2. Results and discussion As illustrated in Fig. 2B and D, the model simulated all the significant effects that were observed in the behavioral data. Overall, the simulated RT was faster for congruent pairs than for incongruent pairs. Like in the behavioral data, the congruity effect was modulated by distance on the (relevant) numerical dimension: it was larger when the numerical distance was small than when the numerical distance was large (Fig. 2B). As also observed in the behavioral data, the opposite pattern was found for distance on the (irrelevant) physical dimension: when physical distance was small, the congruity effect was smaller than when it was large (Fig. 2D). The model activated the incorrect decision on 1.5% of all test trials. These errors were all made in the slowest condition (i.e. incongruent, small numerical distance, large physical distance). It is straightforward to understand how the size congruity effect emerges from the model by examining the examples in Fig. 3. In the case of a congruent stimulus pair like 2 8 (Fig. 3A), the numerical size route will activate the correct right larger decision unit, and so will the physical size route. This results in a fast RT. In the case of an incongruent stimulus pair like 8 2 (Fig. 3B), the physical size route will still activate the correct right larger decision unit. However, the numerical size route will activate the left larger decision, resulting in a slower RT. We now consider why our dual route model simulates the effects of distance. First, because the numerical route in the shared decisions model is based on the model of Verguts et al. (2005), an effect of numerical distance is also observed in the current model. Because the physical route is formally equivalent to the numerical route, it also produces a distance effect. The specific pattern of interactions between the two routes is caused by the non-linear function that transforms the output activation of the decision units into response time. This is similar to the non-linear function that maps drift rate onto RT in the diffusion model of Schwarz and Ischebeck (2003). Intuitively, this can be understood as follows: When distance on the relevant (numerical) dimension is large, there will be a stronger activation induced by the numerical route on the decision units than when the numerical distance is small. Hence, in the former case the irrelevant (physical) dimension will have relatively less influence, and so there will be a smaller congruity effect when numerical distance is large. On the other hand, when distance on the irrelevant dimension is large, there will be a stronger influence on the decision units than when the physical distance is small. It follows that there will be a larger congruity effect when physical distance is large. Importantly, the effects observed in Simulation 1 are caused by the dual route architecture of our shared decisions model and not by specific parameter settings in the model (see Appendix B). 4. Experiment 2 According to the shared decisions account, the size congruity effect originates from the fact that the numerical size route and the physical size route activate shared decision alternatives. These alternatives are set by the task instructions. In Experiment 1 and Simulation 1, the decision alternatives were ‘‘left larger’’ and ‘‘right larger’’ or ‘‘left smaller’’ and ‘‘right smaller’’. In the current experiment, we investigate what happens when the physical and numerical magnitude of all stimuli stays exactly the same, but the decision alternatives are changed by the task instructions. According to the shared representation account, the size congruity effect originates at the level of the shared analog representation of numerical and physical magnitude. If no change is made to the stimulus material, then the representation of numerical and physical magnitude remains unchanged. Hence, the shared representation account does not predict a change in the size congruity effect in this case. In contrast, the size congruity effect in our implementation of the shared decisions account originates from the pattern of weights between the representation of magnitude and the decision. If changing the decision alternatives leads to a change in this pattern of weights during training, then the size congruity effect will also change. To sum up, keeping the stimulus material S. Santens, T. Verguts / Cognition 118 (2011) 97–113 103 Fig. 3. Shared decisions model for solving the number comparison task in a size congruity paradigm. Numerical size is the relevant dimension, physical size the irrelevant dimension. Panel A: congruent stimulus pair. The numerical and the physical size processing routes both activate the same decision (right larger). Panel B: incongruent stimulus pair. Both dimensions activate different decisions: numerical size activates the (correct) left larger decision, but physical size activates the (incorrect) right larger decision. Because numerical size is the relevant dimension, the activation of the left larger decision will be higher than the activation of the right larger decision. the same, but changing the decision alternatives can affect the size congruity effect according to the shared decisions account, but not according to the shared representation account. We test these predictions in Experiment 2. We compare the performance on two different tasks in a size congruity paradigm. The first task is magnitude judgment with one stimulus per trial (e.g., Schwarz & Heinze, 1998; Schwarz & Ischebeck, 2003). The participant has to judge if a number is larger or smaller than the mean numerical size, in this case 5. In this task, the two accounts make the same predictions about the congruity of each stimulus. For example, the stimulus 2 is numerically small, but physically large, so it is incongruent according to the shared representation account. If the physical size of 2 is larger than the standard size, then this stimulus is also incongruent according to the shared decisions account, because the physical dimension activates the ‘‘larger than the standard’’ decision unit, whereas the numerical dimension activates the ‘‘smaller than the standard’’ decision unit (see Fig. 5A). The second task is parity judgment. Here, the participants have to judge whether a stimulus is odd or even. For the shared representation account, 2 is still an incongruent stimulus for the same reason as before. However, at the decision level, the irrelevant dimension no longer causes competition between the two decision alternatives (‘‘odd’’ and ‘‘even’’). The numerical dimension will activate the ‘‘even’’ decision unit for 2, but the physical dimension will not activate any of the two alternatives (see Fig. 5B), because parity cannot be assigned to physical size values. Hence, there will be no congruity effect. In sum, the shared representation account predicts a congruity effect for both the magnitude judgment and the parity judgment task, whereas the shared decisions account only predicts a congruity effect in the magnitude judgment task. 4.1. Method 4.1.1. Participants Sixteen students from Ghent University received course credits to participate in an experiment of 1 h. Three 104 S. Santens, T. Verguts / Cognition 118 (2011) 97–113 Fig. 4. Experiment 2: observed (panels A and C) and simulated (panels B and D) RT. Panels A and B: magnitude judgment task. Panels C and D: parity judgment task (Error bars represent 95% confidence intervals based on the guidelines by Loftus and Masson (1994). Error bars for the model data were too small to depict. To aid visual inspection of the effect sizes, the model data of all simulations are presented on the same scale and so are the behavioral data of all experiments. The range of values on the Y axis is the same for all simulations, whereas it changes between experiments for the behavioral data.) participants were female, 13 were right-handed. Their mean age was 20 years (range: 18–27 years). All participants had normal or corrected to normal vision. They signed an informed consent prior to the experiment. 4.1.2. Stimuli and apparatus The experimental setting was the same as in Experiment 1. Stimuli were now the Arabic numerals 1, 4, 6 and 9. The four different physical sizes of the stimuli were based on the mean values obtained from the calibration procedure in Experiment 1. 4.1.3. Experimental procedure A trial started with the presentation of a small square in the center of the screen that served as the fixation mark. There was a variable time interval before the stimulus appeared, with a duration randomly drawn from a uniform distribution with a minimum of 500 ms and a maximum of 1000 ms. After this interval, the stimulus appeared centrally. Participants had to press the left or right button according to the numerical size of the stimulus (larger or smaller than the mean size, i.e. 5) or according to the parity of the stimulus (odd or even). They were instructed to ignore the physical size of the stimuli. There was a response deadline of 1500 ms. When the response was given or when 1500 ms had elapsed, there was a blank screen for 500 ms until the next trial started. All combinations of numerical and physical sizes were presented (16 stimuli). Each stimulus was repeated four times in each of the 20 blocks in the experiment. This gave a total of 1280 trials. Half of the participants started with the magnitude judgment task for 10 blocks, the other half started with the parity judgment task for 10 blocks. In the magnitude judgment task, smaller than 5 was associated with a left response and larger than 5 with a right response for five consecutive blocks. In the other five blocks, this mapping was reversed. Similarly for the parity judgment task, odd was associated with a left response and even with a right response for five blocks; for the other five blocks this mapping was reversed. The order of tasks and of response mappings was completely counterbalanced between subjects. 4.2. Results and discussion All correct trials entered the RT analysis. Mean RTs per condition were analyzed with an ANOVA with a 2 (task: magnitude judgment or parity judgment) ! 2 (congruity: congruent or incongruent) design. Both variables were treated as within subjects factors. Congruity was defined as in other studies that used one stimulus per trial (e.g., Schwarz & Heinze, 1998; Schwarz & Ischebeck, 2003), similarly for the two tasks. Stimuli were treated as congruent when their numerical and physical sizes were both smaller or both larger than the mean numerical and physical size (e.g., a stimulus with the numerical size 2 and the physical size a) and incongruent otherwise (e.g. a stimulus with the numerical size 2 and the physical size h). There was a main effect of S. Santens, T. Verguts / Cognition 118 (2011) 97–113 105 Fig. 5. (A, B) Neural network model for Experiment 2: the magnitude judgment task (panel a) and the parity judgment task (panel b). (C) Model for the close/far task of Experiment 3. task [F(1, 15) = 13.31; p < .01; g2p ¼ :47]: participants were faster to judge the magnitude of the stimuli than to judge their parity. The main effect of congruity was also significant [F(1, 15) = 7.18; p < .05; g2p ¼ :32]. Importantly, task interacted with congruity [F(1, 15) = 27.36; p < .001; g2p ¼ :65]. Post hoc contrasts revealed that there was a size congruity effect in the magnitude judgment task [F(1, 15) = 30.29; p < .001; g2p ¼ :67] (see Fig. 4A), but not in the parity judgment task [F(1, 15) < 1; g2p ¼ :04] (see Fig. 4B). On average, participants did not respond in time or made an error on 4.3% of all trials. The same ANOVA as for the RT analysis was run for the arcsine transformed error rates. There was no main effect of task [F(1, 15) = 2.17; p = .16; g2p = .13]. There was a significant main effect of congruity [F(1, 15) = 5.31; p < .05; g2p = .26]: Less errors were made on congruent trials than on incongruent trials. The interaction between task and congruity was not significant [F(1, 15) = 1.76; p = .21; g2p ¼ :11]. In sum, these results confirm the predictions derived from the shared decisions account. 5. Simulation 2 5.1. Method The models used in Simulation 2 are presented in Fig. 5A and B. They are similar to the model of Simulation 1, but now only 1 stimulus is present for both dimensions. The models were trained on both magnitude judgment and parity judgment. After training, the performance of the models was tested using the 16 stimuli from Experiment 2 that varied in numerical and physical size. 5.2. Results and discussion After training, the model responded correctly to all stimuli. Simulated RT is plotted in Fig. 4. The model showed a congruity effect for the magnitude judgment task (Fig. 4B), but not for the parity judgment task (Fig. 4D). This is in agreement with the interaction between congruity and task observed in the behavioral data. In the magnitude judgment task, the model exhibits a size congruity effect for the same reason as in Simulation 1. Because a parity judgment on physical sizes is not possible, the model was trained to neither select an odd or an even decision when a physical size from a to i was presented at the input. Therefore, only the numerical size input had an influence on the activation of the decision units and there is no congruity effect in this case. In the model, there is no difference in average RT between the magnitude judgment and the parity judgment task, as was the case in the behavioral data. A possible explanation of this discrepancy could be differential amounts of training on the two tasks in real life, but not 106 S. Santens, T. Verguts / Cognition 118 (2011) 97–113 in the model. However, this difference is beyond the scope of the current study. 6. Experiment 3 In Experiment 2, we showed that the size congruity effect is absent when the decision alternatives required for the task are not automatically activated by the irrelevant dimension. However, caution should be taken in making strong conclusion based on this null effect. It could be argued that in the parity task of Experiment 2, participants were able to completely ignore the physical dimension, because parity does not apply to this dimension. Therefore, in Experiment 3, we further explore the influence of task instructions on the size congruity effect by using a task in which the required decision alternatives can be activated by both the relevant and the irrelevant stimulus dimension. Participants are now instructed to judge the numerical distance of the target digit to the standard as being either close to or far from 5 (close/far task; Santens & Gevers, 2008). In particular, 4 and 6 are considered to be close and 1 and 9 are considered to be far. Importantly, in this task some stimuli will be congruent according to the shared representation account, but incongruent for the shared decisions account and vice versa. According to the shared representation account, congruity is defined by the relation between the numerical and physical magnitude of the stimuli. Stimuli in which numerical and physical size are both large (e.g. 9) or both small (e.g. 1) will always be congruent, independent of the required decision. For the shared decisions account however, congruity originates from competition between the decision alternatives. This competition will be present only if the physical size route automatically activates one of the decisions that are needed to perform the task. For example, in the close/far task, the numerical size route will activate the far (from the standard) decision unit when a stimulus like 1 is presented. If the physical size of 1 is far from the mean physical size, then the physical size can automatically activate the same far (from the standard) decision unit (see Fig. 5C). In other words, if the physical size activates the close/far decision units, then this stimulus is congruent according to the shared decisions account as it does not lead to competition between decisions. However, because numerical size is small and physical size is large for this stimulus, it is incongruent according to the shared representation account (see Fig. 6 for additional examples). In Experiment 3, we will examine whether there is a congruity effect as defined by the shared representation account or as defined by the shared decisions accounts in a close/ far task. 6.1. Method 6.1.1. Participants Seventeen students from Ghent University received course credits to participate in an experiment of 30 min. 14 participants were female, 12 were right-handed. Their mean age was 19 years (range: 18–23 years). All participants had normal or corrected to normal vision. They signed an informed consent prior to the experiment. 6.1.2. Stimuli and apparatus The experimental setting was the same as in Experiments 1 and 2. The stimuli were exactly the same as in Experiment 2. 6.1.3. Experimental procedure Trials were presented as in Experiment 2, only the task differed. Participants now had to press the left or the right button according to the numerical distance (close/far) between the target stimulus and the standard 5. All combinations of numerical and physical sizes were presented. In this way, 16 different stimuli could be constructed. Each stimulus was repeated four times in each of the 10 blocks in the experiment (640 trials). In one half of the experiment, participants were instructed to press the left key for a stimulus that was close to 5 (4 or 6) and the right key for a stimulus that was far from 5 (1 or 9). In the other five blocks, this response mapping was reversed. The order of the response mappings was counterbalanced between subjects. Like in Experiments 1 and 2, no details about the physical sizes were given. Participants were instructed to ignore the differences in physical size. 6.2. Results and discussion Mean RTs of all correctly answered trials entered the RT analysis. In the first ANOVA, congruity was defined according to the shared representation account (i.e. in the same way as in the magnitude judgment task). Stimuli that were both numerically and physically larger (or smaller) than the mean size were regarded as congruent (e.g. 9), stimuli that were not were regarded as incongruent (e.g. 9). For this analysis, the main effect of congruity was not significant [F(1, 16) = 1.23; p = .28; g2p ¼ :07] (see Fig. 7A). In the second ANOVA, congruity was defined according to the shared decisions account: if both the numerical and the physical dimension activated the same decision unit, then the stimulus was regarded as congruent (e.g. 9: numerical and physical dimension activate the far decision unit). If not, it was regarded as incongruent (e.g. 6: numerical dimension activates the close decision unit, but physical dimension activates the far decision unit). For this analysis, the main effect of congruity was significant [F(1, 16) = 5.42; p < .05; g2p ¼ :25] (see Fig. 7C). In contrast with the parity task of Experiment 2, the presence of a congruity effect here implies that the irrelevant physical size was automatically categorized as close or far from the mean size and that this influenced the decision process. This was the case even though the mean physical size had never been presented, nor were any details given about the physical sizes. Note that the null effect of congruity according to the shared representation account cannot be due to a lack of power, as it was computed on exactly the same data as the (significant) congruity effect according to the shared decisions account. On average, participants did not respond in time or made an error on 4,7% of all trials. The ANOVA on the transformed error rates in which congruity was defined according to the shared representation account did not show a significant effect of congruity [F(1, 16) = 2.36; p = .14; g2p ¼ :13], neither did the ANOVA in which congruity was S. Santens, T. Verguts / Cognition 118 (2011) 97–113 107 Fig. 6. Examples of stimuli in Experiments 3 and 4, categorized as congruent or incongruent according to the shared representation and the shared decisions accounts. The four different physical sizes are ordered from left to right and top to bottom within the shared representation panel and within the shared decisions panel. defined according to the shared decisions account [F(1, 16) < 1; g2p ¼ :02]. 7. Simulation 3 7.1. Method The model used for the magnitude judgment task in Simulation 2 was adapted to give the decision close (to the mean size) or far (from the mean size) for both numerical size and physical size. After training, simulated RT was obtained for the same 16 stimuli as used in Experiment 3. 7.2. Results and discussion The model gave the correct answer on all stimuli. Simulated RT is plotted in Fig. 7. The results from the model are in accordance with the behavioral results. When congruity is defined according to the shared representation account, there is no congruity effect (Fig. 7B). However, when defined according to the shared decisions account, there is a clear congruity effect (Fig. 7D). More specifically, simulated RT was faster to stimuli for which the numerical size route and the physical size route activated the same decision unit (close or far), whereas it was slower to stimuli for which the numerical and the physical dimension activated a different decision. 8. Experiment 4 So far, in our behavioral experiments and simulations, we investigated the interaction between numerical size as the relevant dimension and physical size as the irrelevant dimension. In the shared decisions model, the numerical and the physical size route are formally equivalent and therefore interchangeable. The model thus predicts a similar congruity effect when the relevant dimension is physical size and the irrelevant dimension is numerical size. In experiment 4, we investigate whether the novel findings with the close/far task in Experiment 3 can be replicated when the relevant and irrelevant dimensions are interchanged. 108 S. Santens, T. Verguts / Cognition 118 (2011) 97–113 Fig. 7. Experiment 3: observed (panels A and C) and simulated (panels B and D) RT. Panels A and B: congruity defined according to the shared representation account. Panels C and D: congruity defined based on the specific decisions for this task. (Error bars for the model data were too small to depict. To aid visual inspection of the effect sizes, the model data of all simulations are presented on the same scale and so are the behavioral data of all experiments. The range of values on the Y axis is the same for all simulations, whereas it changes between experiments for the behavioral data.) 8.1. Method 8.1.1. Participants Twenty students from Ghent University were paid to participate in an experiment of 30 min. 14 participants were female, 17 were right-handed. Their mean age was 20 years (range: 18–24 years). All participants had normal or corrected to normal vision. They signed an informed consent prior to the experiment. 8.1.2. Stimuli and apparatus The experimental setting was the same as in Experiments 1–3. The stimuli were exactly the same as in Experiments 2 and 3. 8.1.3. Experimental procedure The experimental procedure was very similar to that of Experiment 3. The most important difference was that the physical dimension was now the relevant dimension and the numerical dimension was the irrelevant dimension. Hence, participants had to press the left or the right button according to the physical distance (close/far) between the physical size of the target stimulus and the mean physical size. All combinations of numerical and physical sizes were presented. In this way, 16 different stimuli could be constructed. Each stimulus was repeated four times on each of the 10 blocks in the experiment (640 trials). In one half of the experiment, participants were instructed to press the left key for a stimulus with physical size that was close to the mean physical size (b or g) and the right key for a physical size that was far from the mean physical size (a or h). In the other five blocks, this response mapping was reversed. The order of the response mappings was counterbalanced between subjects. At the beginning of the experiment, all physical sizes, including the mean physical size, were displayed on screen. To further familiarize the participants with the different physical sizes, 25 practice trials were presented. No details about the numerical sizes were given. Participants were instructed to ignore the differences in numerical size. 8.2. Results and discussion The RT analyses were based on the mean RTs per condition. Like in Experiment 3, in the first ANOVA, congruity was defined according to the shared representation account (i.e. in the same way as in the magnitude judgment task). For this analysis, the main effect of congruity was not significant [F(1, 19) < 1; g2p < :01] (see Fig. 8A). In the second ANOVA, congruity was defined according to the shared decisions account. For this analysis, the main effect of congruity was significant [F(1, 19) = 6.24; p < .05; g2p ¼ :25] (see Fig. 8C). To directly test possible differences between Experiment 3 and 4, we analyzed the data from these two experiments with an ANOVA with a 2 (congruity defined according to the shared decisions account: congruent or incongruent) ! 2 (relevant dimension: numerical or physical) design in which congruity was treated as a within subjects factor and relevant dimension as a between subjects factor. Evidently, there was a main effect of congruity S. Santens, T. Verguts / Cognition 118 (2011) 97–113 109 Fig. 8. Experiment 4: observed (panels A and C) and simulated (panels B and D) RT. Panels A and B: congruity defined according to the shared representation account. Panels C and D: congruity defined based on the specific decisions for this task. (Error bars for the model data were too small to depict. To aid visual inspection of the effect sizes, the model data of all simulations are presented on the same scale and so are the behavioral data of all experiments. The range of values on the Y axis is the same for all simulations, whereas it changes between experiments for the behavioral data.). [F(1, 35) = 10.97; p < .01; g2p ¼ :24]. There was also a main effect of relevant dimension [F(1, 35) = 27.66; p < .001; g2p ¼ :44]. Participants responded faster when the numerical dimension was relevant than when the physical dimension. Importantly, there was no interaction between congruity and relevant dimension F(1, 35) < 1; g2p < :01]. On average, participants did not respond in time or made an error on 8.2% of all trials. The same ANOVAs as for the RT analyses were run for the arcsine transformed error rates in. The ANOVA in which congruity was defined according to the shared representation account did not show a significant effect of congruity [F(1, 19) = 1.51; p = .23; g2p ¼ :07]. There was a main effect of congruity when it was defined according to the shared decisions account [F(1, 19) = 9.44; p < .01; g2p = .33]. 9. Simulation 4 9.1. Method The model simulation was exactly the same as in Simulation 3, except that now the activity from the numerical dimension was attenuated relative to the activity from the physical dimension. 9.2. Results and discussion The model gave the correct response to all stimuli. Simulated RT is plotted in Fig. 8. The results from the model are in accordance with the behavioral results and very similar to the results of Simulation 3. In particular, simulated RT was sensitive to congruity as defined by the shared decision account rather than the shared representation account. 10. General discussion In this study, we compared the shared representation and the shared decisions accounts of the size congruity effect. Both accounts were hitherto underspecified. In addition, we have argued that the shared representation account suffers from a conceptual indeterminacy problem. The shared decisions account, on the other hand, can be considered as belonging to a broader theoretical framework, explaining congruity effects in terms of dual route processing. We specified this account in a dual route model, based on the number comparison mechanism in the model of Verguts et al. (2005). We simulated the congruity effect and its interactions with distance on the relevant and the irrelevant dimension (Experiment 1, Simulation 1). In Experiment 2, using two different tasks but identical stimuli across tasks, the size congruity effect was shown to be dependent on the task instructions in a way predicted by the shared decisions account. In Experiment 3, it was shown that the task not only modulates the presence of a size congruity effect, but that it also predicts specific congruity effects different from those in a number comparison task. In Experiment 4, we showed that interchanging the relevant and the irrelevant dimensions does not necessarily have an influence on the size 110 S. Santens, T. Verguts / Cognition 118 (2011) 97–113 congruity effect. Indeed, whereas there was an overall difference in RTs between Experiments 3 and 4, there was no interaction with size congruity. Our dual route model, as an instantiation of the shared decisions account, was able to simulate all of these effects. In contrast, it is difficult to see how the results from Experiments 2–4 would be explained by an account in which the size congruity effect originates from a level where numerical and physical size would be jointly represented in an analog format (shared representation account). Our computational model can be related to the coalescence model of Schwarz and Ischebeck (2003). Interestingly, this diffusion model approach accurately describes RT and accuracy in a standard size congruity paradigm, and has the potential to also investigate more detailed aspects of RT and accuracy distributions. However, its current formulation does not distinguish between the different accounts on how numerical and physical size interact. As the authors state themselves ‘‘. . . there is nothing in the formalism of the coalescence model per se that could distinguish between these accounts in terms of an early versus late interaction’’ (p. 520). The computational model presented here has a complementary advantage. Whereas it incorporates a simpler way of simulating RTs and errors, it explicitly models how physical and numerical size interact. In doing so, it makes only the necessary assumptions (see Appendix B). We believe that joining the individual qualities of these two approaches may provide a more realistic model of how quantities are compared; we leave that project for a follow-up study. An open question is to what extent the size congruity effect can be observed when numerical size is paired with quantitative dimensions other than physical size. Size congruity effects have already been observed between numerical size and quantities such as luminance (Cohen Kadosh & Henik, 2006; Cohen Kadosh et al., 2008; Pinel et al., 2004), line length (Dormal & Pesenti, 2007, 2009), duration (Dormal, Andres, & Pesenti, 2008) and dot size (Gebuis, Cohen Kadosh, de Haan, & Henik, 2009). In each case, the dual route model presented here predicts that the effects will originate from the decision level and not from the representational level. However, this does not imply that each pair of dimensions should exhibit equal size congruity effects. Some dimensions may be more discriminable (Pansky & Algom, 1999), and some pairs of dimensions might have more similar decisions. For example, in a comparison task, numerical and physical size may have more similar decisions (smaller and larger) than numerical size and brightness (smaller versus less bright and larger versus more bright). In general, we predict that each quantitative dimension can interfere with any other to the extent that the irrelevant dimension is effective in automatically activating the decision units that are used for the task at hand. We have shown that the shared decisions model presented in this study can account for both well-known and novel congruity effects. However, this does not exclude the possibility that a shared representation for different quantities exists. Strong evidence in favor of a shared representation would be single neurons that show selectivity for quantity in different dimensions. Several authors have put forward the hypothesis that the neurons that are initially recruited to represent space and time are used later in development to represent number as well (Bueti & Walsh, 2009; Cohen Kadosh & Walsh, 2008; Cohen Kadosh et al., 2008). In search of such neurons, Tudusciuc and Nieder (2007) presented dot patterns and line lengths to rhesus monkeys while recording neural activity from the intraparietal sulcus (IPS), a site where neurons selective to number have been observed (Nieder & Miller, 2004). Only about 3–4% of the 400 neurons in IPS they recorded from had both a preferred numerosity and a preferred line length. These cells would thus be good candidates to represent quantity independent of the specific dimension. However, if these neurons represent a common metric, then their preferred numerosity should correlate with their preferred line length. This was not the case. Hence, not even these few neurons can be said to have a common metric for number and length. It was, however, the case that number-selective and length-selective neurons were spatially intermixed in the IPS. One possibility is that (a subset of) neurons in IPS are naturally ‘‘biased’’ for coding quantitative dimensions, and that different neurons become recruited for different dimensions depending on which dimensions are important for the organism. This account remains however to be tested. Appendix A. Calibration of the stimuli for Experiment 1 On the first day of Experiment 1, participants performed a calibration procedure. In this procedure, the physical sizes of the stimuli were adjusted in order to minimize the response time (RT) differences between comparing numerical sizes 1 and 2 and physical sizes a and b, between comparing numerical sizes 2 and 7 and physical sizes b and g and between comparing numerical sizes 7 and 8 and physical sizes g and h. The values for a, b, g, and h were defined as factors with which the height of the digits in the first block was multiplied. The experiment started with a numerical comparison task, in which this factor was fixed at 1 for all stimuli. This corresponded to a height of 7.4 visual degrees. 20 repetitions of the stimulus pairs 1–2, 2–7, and 7–8 (i.e., 60 trials) were presented. The order of the digits in a pair was randomized. At the end of these 60 trials, median RTs for each pair were calculated and participants were informed that now the physical comparison task had to be performed. Another 60 trials were presented, but now consisting of 20 repetitions of each of the physical size pairs a–b, b–g, and g–h. For the first 60 trials of physical comparison, size a was set to 0.3, b to 0.55, g to 1.45 and h to 1.70 times the height of the stimuli in the numerical comparison task. The arbitrary symbols ‘‘#’’, ‘‘@’’, ‘‘&’’ and ‘‘}’’ were randomly assigned to these physical heights. As in the first 60 trials, participants had to press on the side of the smaller or larger stimulus, but now based on its physical size. After a block of 60 + 60 = 120 trials, the mean differences in RT between pair 1–2 and pair a–b (DRT1), between pair 2–7 and pair b–g (DRT2) and between pair 7–8 and pair g–h (DRT3) were calculated. If this difference was positive (i.e., a higher RT for the numerical pair than for the physical pair), then the distance between the two physical values was decreased to make them less S. Santens, T. Verguts / Cognition 118 (2011) 97–113 discriminable. If the difference was negative, then the distance between the physical values was increased. More RT 2 specifically, D1000 was first added to a and b and subtracted RT 1 RT 3 from g and h. Then, D500 was added to a and D500 was subtracted from h. Ten such blocks of 120 trials were presented in the calibration phase. In this way, the values for a, b, g and h were determined for each participant separately and used in the size congruity task on the second day of the experiment. Appendix B. Neural network model implementation B.1. Simulation 1 The model in Simulation 1 implements a size congruity task according to the shared decisions account. It is based on the Verguts et al. (2005) number comparison model. For simplicity, we did not implement input or output layers in the model, only the representation and decision layers are included (for a similar approach, see Van Opstal et al., 2008). The physical dimension is implemented by adding two extra ‘‘representation layers’’ (one for the right stimulus and one for the left; see Fig. 2) with a similar structure as the number line layers in the original model of Verguts et al. (2005). The numerical size representation layers and the physical size representation layers are connected to the same decision units. These units are labeled with the decisions left larger and right larger. Each representation layer has 9 units. Numerical size is represented by an activation pattern that corresponds to a Gaussianlike function (Verguts & Fias, 2004). Physical size is represented identically; it is assumed that there are 9 sizes (a–i) and the distance between two successive sizes is equal to the distance between two successive numerical sizes. The net input activation to the decision units was transferred using a log-sigmoid function. The simulated RT of the network was calculated as 1 ðjo1 $ o2 j þ cÞ with o1 and o2 being the activation of the decision units and c being a small constant, equal to 1 for all simulations. Using the delta rule, the network was trained separately to give the correct response for the numerical comparison task and the physical comparison task, respectively. When trained on one dimension, the units of the representation layer of the other dimension were set to zero. Training was performed using a gradient descent with momentum learning function. The learning rate was set to 5. During training, weights were adjusted to minimize the average squared error (mse) between the values of the output units and the target values for these units (i.e. [0 1] or [1 0]). The model was trained until this mse was smaller than 0.001. Behavioral performance in Experiment 1 was simulated by testing the trained model with the numerical sizes 1, 2, 7 and 8, and the physical sizes a, b, g and h. During testing, the activation of the representation layer units on the irrelevant dimension was multiplied by a parameter H with the value 0.15; this is a proxy to biasing the relevant dimension via task demand units implemented in earlier 111 models (e.g., Botvinick, Braver, Barch, Carter, & Cohen, 2001; Cohen et al., 1990). It is also similar to the H parameter in the coalescence model by Schwarz and Ischebeck (2003). In their diffusion model approach, this parameter reflects ‘‘. . .the degree to which the drift rate component of the irrelevant attribute enters into the overall drift rate.’’ Training and testing of the model was repeated as many times as there were participants in the corresponding behavioral experiment. The means of the simulated RTs over these replications are presented in the results section. The size congruity effect originates in the model from the fact that, although numerical and physical sizes are represented separately, the decision units are shared for the two dimensions. The H parameter biases the relevant dimension over the irrelevant with respect to the activity that reaches the decision units. The distance effects on the relevant and the irrelevant dimension are caused by the pattern of weights between the representations of numerical and physical size and the decision units, just like in the model of Verguts et al. (2005; also see Van Opstal et al., 2008). The specific interaction pattern between distance on the relevant and irrelevant dimension is caused by the non-linear function transforming the activation of the decision units into response time. We tested the effect of varying parameter values on the performance of the model. The gradient descent with momentum learning function and the learning rate were chosen for computational efficiency only. Increasing the learning to e.g. 50 and choosing an even more efficient learning function (gradient descent with momentum and adaptive learning rate) resulted in very fast learning (the performance goal was reached in less than 100 iterations), but caused only minor quantitative and no qualitative changes in the simulated RT that the model produced. The same was true when learning was made very slow by setting the learning rate to 5 and choosing a less efficient learning function (gradient descent without momentum nor adaptive learning rate). Decreasing the performance goal (e.g. to a value as low as 0.0001) only had a minor influence on the behavior of the model, but naturally increased the training time. Increasing or decreasing the width of the activation spread in the numerical and physical size representation layers also had only a minor influence and did not change the results qualitatively. This is because the comparison mechanism in our model is based on the pattern of weights between the representation and the decision units and not on the representation itself (for a detailed explanation, see Van Opstal et al., 2008; Verguts et al., 2005). Finally, we tested the influence of varying the value of the H parameter. Decreasing this value (e.g. to 0.05) led to a diminished size congruity effect. Increasing its value led to a bad performance of the model. For values of 0.30 or higher, the model made many errors, especially in the conditions that were the slowest in the original simulation. B.2. Simulation 2 Only minor adjustments were made to the model of Simulation 1 (see Fig. 5A). There was now only one representation layer for the numerical dimension and one for 112 S. Santens, T. Verguts / Cognition 118 (2011) 97–113 the physical dimension, as only one stimulus was presented in each trial. The decision units were now labeled smaller (than the mean size) and larger (than the mean size). For the parity judgment task, the model architecture was similar, but the decision units were labeled odd and even (see Fig. 5B). The models were trained separately for numerical size input and physical size input. Because physical sizes cannot be categorized as odd or even, the model was trained to have zero activation for both decision units when physical sizes were presented in the parity task. B.3. Simulation 3 The model used in Simulation 3 is identical to that in Simulation 2 (see Fig. 5), except for the decisions used to perform the task. The decision units were now labeled close (to the mean size) and far (from the mean size). The model was trained separately on numerical size and physical size input. Similarly to the behavioral experiment, the model was trained to answer close when numerical sizes 4 or 6 were presented and far when numerical sizes 1 or 9 were presented. For physical size input, the model was trained to answer close for physical sizes b and g and far for sizes a and h. B.4. Simulation 4 The only difference with Simulation 3 is that here, the physical dimension was treated as the relevant dimension and the numerical dimension as the irrelevant. References Ansari, D., Fugelsang, J. A., Dhital, B., & Venkatraman, V. (2006). Dissociating response conflict from numerical magnitude processing in the brain: An event-related fMRI study. NeuroImage, 32, 799–805. Algom, D., Dekel, A., & Pansky, A. (1996). The perception of number from the separability of the stimulus: The Stroop effect revisited. Memory & Cognition, 24, 557–572. Besner, D., & Coltheart, M. (1979). Ideographic and alphabetic processing in skilled reading of English. Neuropsychologia, 17, 467–472. Botvinick, M. M., Braver, T. S., Barch, D. M., Carter, C. S., & Cohen, J. D. (2001). Conflict monitoring and cognitive control. Psychological Review, 108, 624–652. Bueti, D., & Walsh, V. (2009). The parietal cortex and the representation of time, space, number and other magnitudes. Philosophical Transactions of the Royal Society B – Biological Sciences, 364, 1831–1840. Cantlon, J. F., Platt, M. L., & Brannon, E. M. (2009). Beyond the number domain. Trends in Cognitive Sciences, 13, 83–91. Chiou, R. Y.-C., Chang, E. C., Tzeng, O. J.-L., & Wu, D. H. (2009). The common magnitude code underlying numerical and size processing for action but not for perception. Experimental Brain Research, 194, 553–562. Cohen, J. D., Dunbar, K., & McClelland, J. L. (1990). On the control of automatic processes – A parallel distributed processing account of the Stroop effect. Psychological Review, 97, 332–361. Cohen, J. D., Servan-Schreiber, D., & McClelland, J. L. (1992). A parallel distributed processing approach to automaticity. The American Journal of Psychology, 105, 239–269. Cohen Kadosh, R., Cohen Kadosh, K., & Henik, A. (2008). When brightness counts: The neuronal correlate of numerical-luminance interference. Cerebral Cortex, 18, 337–343. Cohen Kadosh, R., Cohen Kadosh, K., Linden, D. E. J., Gevers, W., Berger, A., & Henik, A. (2007). The brain locus of interaction between number and size: A combined functional magnetic resonance imaging and event-related potential study. Journal of Cognitive Neuroscience, 19, 957–970. Cohen Kadosh, R., & Henik, A. (2006). A common representation for semantic and physical properties. Experimental Psychology, 53, 87–94. Cohen Kadosh, R., Lammertyn, J., & Izard, V. (2008). Are numbers special? An overview of chronometric, neuroimaging, developmental and comparative studies of magnitude representation. Progress in Neurobiology, 84, 132–147. Cohen Kadosh, R., & Walsh, V. (2008). From magnitude to natural numbers: A developmental neurocognitive perspective. Behavioral and Brain Sciences, 31, 647–648. Coles, M. G. H. (1989). Modern mind-brain reading: Psychophysiology, physiology, and cognition. Pscyhophysiology, 26, 251–269. Dormal, V., Andres, M., & Pesenti, M. (2008). Dissociation of numerosity and duration processing in the left intraparietal sulcus: A transcranial magnetic stimulation study. Cortex, 44, 462–469. Dormal, V., & Pesenti, M. (2007). Numerosity-length interference – A Stroop experiment. Experimental Psychology, 54, 289–297. Dormal, V., & Pesenti, M. (2009). Common and specific contributions of the intraparietal sulci to numerosity and length processing. Human Brain Mapping, 30, 2466–2476. Gebuis, T., Cohen Kadosh, R., de Haan, E., & Henik, A. (2009). Automatic quantity processing in 5-year olds and adults. Cognitive Processing, 10, 133–142. Gevers, W., Ratinckx, E., De Baene, W., & Fias, W. (2006a). Further evidence that the SNARC effect is processed along a dual-route architecture – Evidence from the lateralized readiness potential. Experimental Psychology, 53, 58–68. Gevers, W., Verguts, T., Reynvoet, B., Caessens, B., & Fias, W. (2006b). Numbers and space: A computational model of the SNARC effect. Journal of Experimental Psychology – Human Perception and Performance, 32, 32–44. Hackley, S. A., & Miller, J. (1995). Response complexity and precue interval effects on the lateralized readiness potential. Psychophysiology, 32, 230–241. Henik, A., & Tzelgov, J. (1982). Is 3 greater than 5 – The relation between physical and semantic size in comparison tasks. Memory & Cognition, 10, 389–395. Kaufmann, L., Koppelstaetter, F., Delazer, M., Siedentopf, C., Rhomberg, P., Golaszewski, S., et al. (2005). Neural correlates of distance and congruity effects in a numerical Stroop task: An event-related fMRI study. NeuroImage, 25, 888–898. Masaki, H., Wild-Wall, N., Sangals, J., & Sommer, W. (2004). The functional locus of the lateralized readiness potential. Psychophysiology, 41, 220–230. Meck, W. H., & Church, R. M. (1983). A mode control model of counting and timing processes. Journal of Experimental Psychology: Animal Behavior Processes, 9, 320–334. Nieder, A., & Miller, E. K. (2004). A parieto-frontal network for visual numerical information in the monkey. Proceedings of the National Academy of Sciences of the United States of America, 101, 7457–7462. Loftus, G. R., & Masson, M. E. J. (1994). Using confidence-intervals in within-subject designs. Psychonomic Bulletin & Review, 1, 476–490. Pansky, A., & Algom, D. (1999). Stroop and Garner effects in comparative judgment of numerals: The role of attention. Journal of Experimental Psychology – Human Perception and Performance, 25, 39–58. Pinel, P., Piazza, M., Le Bihan, D., & Dehaene, S. (2004). Distributed and overlapping cerebral representations of number, size, and luminance during comparative judgments. Neuron, 41, 983–993. Ratcliff, R. (1978). A theory of memory retrieval. Psychological Review, 85, 59–108. Santens, S., & Gevers, W. (2008). The SNARC effect does not imply a mental number line. Cognition, 108, 263–270. Schwarz, W., & Heinze, H. J. (1998). On the interaction of numerical and size information in digit comparison: A behavioral and event-related potential study. Neuropsychologia, 36, 1167–1179. Schwarz, W., & Ischebeck, A. (2003). On the relative speed account of number-size interference in comparative judgment of numerals. Journal of Experimental Psychology – Human Perception and Performance, 29, 507–522. Stevens, M., Lammertyn, J., Verbruggen, F., & Vandierendonck, A. (2006). Tscope: A C library for programming cognitive experiments on the MS Windows platform. Behavior Research Methods, 38, 280–286. Stroop, J. R. (1935). Studies of interference in serial verbal reactions. Journal of Experimental Psychology, 18, 643–662. Szücs, D., & Soltész, F. (2008). The interaction of task-relevant and taskirrelevant stimulus features in the number/size congruency paradigm: An ERP study. Brain Research, 1190, 143–158. Tudusciuc, O., & Nieder, A. (2007). Neural population coding of continuous and discrete quantity in the primate posterior parietal cortex. Proceedings of the National Academy of Sciences of the United States of America, 104, 14513–14518. S. Santens, T. Verguts / Cognition 118 (2011) 97–113 Usher, M., & McClelland, J. L. (2001). The time course of perceptual choice: The leaky, competing accumulator model. Psychological Review, 108, 550–592. Van Opstal, F., Gevers, W., De Moor, W., & Verguts, T. (2008). Dissecting the symbolic distance effect: Comparison and priming effects in numerical and nonnumerical orders. Psychonomic Bulletin & Review, 15, 419–425. Verguts, T., & Fias, W. (2004). Representation of number in animals and humans: A neural model. Journal of Cognitive Neuroscience, 16, 1493–1504. Verguts, T., & Fias, W. (2008). Symbolic and nonsymbolic pathways of number processing. Philosophical Psychology, 21, 539–554. 113 Verguts, T., Fias, W., & Stevens, M. (2005). A model of exact small-number representation. Psychonomic Bulletin & Review, 12, 66–80. Walsh, V. (2003). A theory of magnitude: common cortical metrics of time, space and quantity. Trends in Cognitive Sciences, 7, 483–488. Xuan, B., Zhang, D., He, S., & Chen, X. (2007). Larger stimuli are judged to last longer. Journal of Vision, 7, 1–5. Yeung, N., Botvinick, M. M., & Cohen, J. D. (2004). The neural basis of error detection: Conflict monitoring and the error-related negativity. Psychological Review, 111, 931–959. Zorzi, M., & Umiltà, C. (1995). A computational model of the Simon effect. Psychological Research – Psychologische Forschung, 58, 193–205.
© Copyright 2026 Paperzz