Numerical cognition has long been thought to represent a prime example of an abstract propositional symbolic system with no obvious reference to the outer world. For example, number names such as ‘one’, ‘two’ and so on do not bear any obvious association to the referenced cardinal meaning. Recent evidence, however, implies that numbers and mental arithmetic bear numerous associations with space and time (Fischer & Knops, 2014; Hubbard, Piazza, Pinel, & Dehaene, 2005). Beyond using space to represent numerical magnitude along a spatially oriented mental number line (MNL), it has been hypothesized that humans reuse their spatial prowess and the underlying neural resources to master numerical operations (Dehaene & Cohen, 2007; Knops, Thirion, Hubbard, Michel, & Dehaene, 2009). The cortical mechanisms in parietal cortex that serve for the transformation of spatial coordinates in different referent frames (Beck, Latham, & Pouget, 2011; Pouget, Deneve, & Duhamel, 2002) may play a pivotal role in this context. In particular, neural populations in posterior superior parietal cortex encode object location with respect to different body parts and references. For example, during the planning of eye movements these circuits compute the headcentred position of a given object by adding the activity of neural populations that encode eyecentred coordinates and eye position (Beck et al., 2011; Pouget et al., 2002). Hence, the cortical algorithms for executing simple additions are present in parietal cortex. Basic mathematical operations might coopt these circuits by providing numerical input instead of coordinate information (Dehaene & Cohen, 2007; Hubbard et al., 2005; Knops, Thirion, et al., 2009).
This process is not biasfree. When approximating the outcome of simple addition or subtraction problems humans are likely to provide a biased response that deviates systematically from the correct outcome. The results of addition problems tend to be overestimated while the results for subtraction problems are underestimated. This cognitive bias is referred to as operational momentum (OM; McCrink, Dehaene, & DehaeneLambertz, 2007) due to its resemblance to a systematic bias in the estimation of the point in space where a moving object disappeared (Hubbard, 2005, 2015). OM has initially been observed with nonsymbolic operations where participants estimated the number of visual objects (dots) that would result from adding two sets of dots or subtracting one set from the other (McCrink et al., 2007). However, OM has also been found in symbolic notation (Knops, Dehaene, Berteletti, & Zorzi, 2014; Knops, Viarouge, & Dehaene, 2009) which has been interpreted as evidence for a common underlying mechanism. OM has also been observed with paradigms that require translating the cognitively generated numerical estimate into a position on a labeled line (Pinhas & Fischer, 2008), or actively producing the outcome via a dot generating manual device (Lindemann & Tira, 2011). In the labeled line task, participants indicated the position of the outcome of visually presented addition and subtraction problems on a line that was labeled with zero on the left and 10 on the right. Crucially, participants misplaced addition outcomes to the right and subtraction outcomes to the left, compared to baseline that required indicating the numbers’ positions without preceding arithmetic operation. When actively producing dot patterns corresponding to the outcome of multidigit problems, participants produced relatively larger estimates for addition problems as compared to subtraction problems with identical correct outcome (Lindemann & Tira, 2011). Together, this implies that mental arithmetic is subject to systematic biases that may have their origin in the application of spatial coordinate transformation mechanisms to numerical quantity information that can be conceived of as positions on the MNL.
More evidence for spatial contributions to mental arithmetic comes from a recent study reporting systematic interference between arm or eye movements and mental arithmetic (Wiemers, Bekkering, & Lindemann, 2014). Addition performance was impaired when participants moved their arms or eyes downward while subtraction was affected with upward movements of the arm or eye. In horizontal plane only arm but not eye movements to the left interfered with addition while arm movements to the right interfered with subtraction. During simple addition and subtraction tasks, Margethis and colleagues found systematic deviations of manual mouse pointer trajectories from the ideal path such that addition and subtraction trajectories deviated to the right and left from the ideal paths, respectively (Marghetis, Núñez, & Bergen, 2014). Interestingly, this effect was observed over and above confounding the response position with numerical magnitude (Pinhas & Fischer, 2008; Pinhas, Shaki, & Fischer, 2014) because the correct outcomes were presented both to the upper left or right relative to starting position.
Despite the growing number of studies demonstrating OM in different settings, the underlying mechanisms of the OM effect are currently debated. Three major hypotheses vie with one another. First, it has been suggested that the OM effect reflects the outcome of a simple heuristic that would associate different arithmetic operations with expectations concerning the numerical relationship between the outcome and the operands (McCrink et al., 2007). For addition, this heuristic would predict that the result should be larger than either of both operands. For subtraction, the heuristic would predict outcomes that are smaller than the first operand. The same heuristic would hold for other arithmetic operations such as multiplication and subtraction. In particular, the observed OM in 9montholds supports this account because it is unlikely that at that age infants have developed a spatial representation of numerical magnitude (McCrink & Wynn, 2009). A second hypothesis assumes that OM results from the flawed logarithmic compression and decompression into a linear metric during the arithmetic process (Chen & Verguts, 2012). The mental number line is assumed to be logarithmically compressed (Dehaene, 2001; Nieder & Dehaene, 2009). The compression and decompression processes needed to transduce between linear and logarithmic scales may be flawed. In the extreme variant of the compression hypothesis, addition and subtraction operate on compressed values, that is, the decompression fails entirely. This would result in massive over and underestimations since the sum of the logs is the log of a product (i.e. log(a) + log(b) = log(a × b)). An equivalent underestimation would result for subtraction since the difference of the logs is the log of a quotient (i.e. log(a) – log(b) = log(a/b)).
Finally, spatial accounts have been proposed to account for OM. According to the spatial competition hypothesis (Pinhas & Fischer, 2008; Pinhas, Shaki, & Fischer, 2015), OM is the result of the “competing spatial biases invoked by the operands, the operation sign, and the result of an arithmetic problem” (p. 997; Pinhas et al., 2015). Operands and results activate their respective positions on the mental number line, that “compete for responses” (p. 413; Pinhas & Fischer, 2008). For subtractions, for example, the result of a given problem may be located between the operands (e.g. 7  2 = 5) or to the left of both operands (e.g. 7  4 = 3). Compared with problems involving zero as a second operand, these competing biases mitigate the observed bias towards the result which is more pronounced when the second operand does not additionally compete for responses. Consequently, Pinhas and Fischer (2008) observed the largest OM bias with zero problems. Recently, these authors observed an inverse OM when reversing the line labels such that the right end of the line was labeled with 0 and the left end with 10 (Pinhas et al., 2015), providing support for their spatial competition bias and underlining the observation that numbertospace mappings are highly flexible (Bächtold, Baumuller, & Brugger, 1998). According to a second spatial account, the OM reflects systematic biases from the deployment of the coordinate transformation system in parietal cortex that also mediates attentional shifts in space (Knops et al., 2014; Knops, Thirion, et al., 2009; Knops, Viarouge, et al., 2009; Knops, Zitzmann, & McCrink, 2013). According to this approach, approximate mental arithmetic is mediated by a dynamic interaction between positional codes on the MNL and an attentional system that shifts the spatial focus to the left or right. At the neural level this may be instantiated in the functional interactions between areas along the intraparietal sulcus and posterior, superior parietal areas. The idea is that a parietal circuit that has been proposed to combine retinal and eye position information via vector addition in order to compute positions in space may be recycled to implement mental arithmetic. The resulting positional information can be used to guide eye or hand movements and has been proposed to be the base for shifts of spatial attention. This places mental arithmetic in the realm of dynamic updating processes of spatial coordinates in parietal cortex and stipulates that the efficiency of this system is linked with arithmetic performance. Due to the approximate nature of this process the shifts may ‘overshoot’, leading to over and underestimation in addition and subtraction, respectively. Interestingly, the latter approach suggests a functional coupling between eye movements and arithmetic. A recent study provided confirmatory evidence for this notion (Klein, Huber, Nuerk, & Moeller, 2014). Participants’ eye movements after the first saccade were observed to move to the right during addition problems and to the left in subtraction problems when asked to indicate the location of the result on a labeled line (but see Hartmann, Mast, & Fischer, 2015; Klein et al., 2014). A second implication from this account is that OM is not limited to addition and subtraction but generalizes to basic arithmetic transformations with natural numbers such as multiplication and division as long as they require quantity manipulations that are associated with attentional shifts along the MNL. Operations that lead to larger outcomes would be associated with rightward shifts of attention, while the opposite should hold for operations that lead to smaller outcomes. Katz and Knops (2014) investigated the OM effect in the context of multiplication and division. However, as opposed to simple addition and subtraction problems, finding the solution of simple symbolic multiplication problems is often conceptualized as verbally mediated recall from long term memory (e.g. Campbell & Xue, 2001), mixed with idiosyncratic short cuts (e.g. retrieving 9 × 7 = 63 by subtracting 7 from 70). Nevertheless, approximate estimates of the outcome might involve spatial transformations, protecting us against accepting grossly wrong solutions. For example, knowing that 32 × 8 must be a threedigit number helps excluding 40 as a possible outcome. Despite this possible role of spatial transformations in multiplication and division, neither for the standard set of multiplication tables (i.e. between 1 × 1 = 1 and 9 × 9 = 81), nor for twodigit × onedigit problems (e.g. 14 × 3 = 42) an OM effect was observed (Katz & Knops, 2014). However, for the corresponding nonsymbolic problems where the quantities were presented as dot patterns a regular OM effect was observed. One crucial difference between the procedure adopted by Katz and Knops (2014) and previous studies (Knops, Viarouge, et al., 2009) was that the correct solution was presented as one of five response alternatives, potentially encouraging the engagement in exact calculation and direct retrieval rather than approximating the outcome. This may have reduced the opportunity to detect any systematic biases due to attentional shifts that accompany approximate quantity manipulations.
The aim of the current study was twofold. First, we aimed at testing the presence of OM in multiplication and division by eliminating the presence of the correct outcome amongst the response alternatives. By encouraging participants to approximate even in the symbolic notation we aimed at increasing sensitivity to detect any systematic biases during multiplication and division with Arabic digits. Second, engaging participants in both an OM task and a Posner paradigm allowed us to test whether potential OM biases actually correlate with attentional measures. According to the above theoretical accounts of the OM we can break this question down into four aspects. Do attentional parameters correlate with (a) a heuristic according to which multiplication leads to larger outcomes and division to smaller outcomes, (b) flawed decompression, (c) competing spatial biases by the operands, the results or the outcome, or (d) attentional shifts along the mental number line? According to the heuristics account and the compressiondecompression approach, no correlation with attentional measures would be predicted. Among the spatial accounts, only the attentional explanation predicts a correlation between attentional parameters and OM effect. No such correlation is predicted by the competing spatial biases account.
Methods [TOP]
Participants [TOP]
Participants (n = 18; 1974 yearsold, M = 36.3, SD = 19; 14 female, 4 male) were recruited from the general population (both student and nonstudent) using a departmental database. Written informed consent was obtained from all participants prior to participation. Students received course credit; no other reimbursement was given.
Procedure and Materials [TOP]
The study consisted of two experiments; a calculation task involving symbolic and nonsymbolic multiplication and division problems, and a variant of the Posner task to test different aspects of visuospatial attention (orienting/selection and reorienting/executive attention).
Calculation Task [TOP]
The calculation task was created and presented using OpenSesame (Mathôt, Schreij, & Theeuwes, 2012). Participants were given written instructions and then performed 24 (12 symbolic, 12 nonsymbolic; intermixed) practice calculation trials. Symbolic (Arabic digits) and nonsymbolic (dotarrays) multiplication and division problems were presented horizontally (separated by ‘x’ or ‘÷’) to reduce working memory demands, for 2000ms, followed by seven response choices presented in a circle until a response was made or for a maximum of 6000ms (Figure 1). Compared to previous studies (Katz & Knops, 2014), we reduced presentation time of the operands to minimize propensity of adopting a counting strategy. Responses were made with a mouseclick on the chosen value or dotarray.
Nonsymbolic stimuli were created using MATLAB and the Psychophysics Toolbox extension (Brainard, 1997; Pelli, 1997). To prevent that participants relied on nonnumeric stimulus features such as density, occupied area or individual dot size we decorrelated quantity from visual parameters (area subtended, average dotsize) in each presented set of response alternatives using the method described by Gebuis and Reynvoet (2011). This method resulted in trialspecific response sets (7 dotarrays) with no correlations between quantity, average dot size or area subtended (.2 < r < .2 respectively).
To catch random responding, symbolic and nonsymbolic control trials (16 ÷ 1, 16 x 1) were intermixed with calculation trials. Participants whose performance deviated more than 3 SD from the group mean or with symbolic control problem accuracy below 50% (chance = 14.3%) were excluded from the study.
To control for response choice magnitude effects, the same result values (with a random jitter in symbolic trials) or quantities (nonsymbolic trials) were presented for multiplication and division; this resulted in different operands for multiplication versus division. We created a geometric series of 11 values ranging from 1/3 to three times the correct value (Knops, Viarouge, et al., 2009; Knops et al., 2013) for nonsymbolic notation. For problems presented in symbolic notation the geometric series spanned values ranging from ½ times to two times the correct value. Arithmetic problems with symbolic range are reported in Appendix A. Because previous findings suggested that presenting the exact correct value may have made responses too accurate to detect symbolic response bias (Katz & Knops, 2014), symbolic response choices were jittered so that the correct answer was never presented (Knops, Viarouge, et al., 2009; Knops et al., 2013). To achieve this, all results were jittered either up or down by a random value which fell within plus or minus half the numerical interval between the correct result and the first deviant above or below it. The random value was drawn from a flat distribution on a logarithmic scale with a mean value of zero and was fixed for a given trial.
Participants completed 576 calculation trials (144 per condition) and 120 control trials (16 ÷ 1, 16 x 1; 30 per condition). To prevent the correct result always being the median value (i.e. 4^{th} response value rank of 7 choices), we varied the range of response alternatives and presented only seven out of the eleven response alternatives for each problem. The seven response alternatives corresponded to the smallest (low), largest (high) or middle (middle) range of response alternatives (Figure A.1). This means that the response alternative which is closest to the correct outcome and upon which the response alternatives were centred changed its ordinal rank for the different ranges. That is, for the smallest range, it fell close to the upper end (rank six out of seven), while it was located close to the lowest end for the highest range (rank two out of seven). In calculation trials, each problem was presented six times (two per range). At the end of the calculation task, participants were asked to describe how they solved the problems (“Please describe how you solved the problems. Which strategies did you use to decide for one of the response alternatives?”). Responses were typed in a blank document and no word limit was given. Strategy and word count were extracted from the responses.
Attention Task [TOP]
Attention was assessed using an endogenous Posner cueing task (Posner, 1980), created and presented with MATLAB and the Psychophysics Toolbox (PTB) extension (Brainard, 1997; Pelli, 1997). Participants were seated approximately 60cm from the display and instructed to fixate on a red fixation dot (0.6°), which was presented at the center of the screen, flanked by a dark grey square box on either side (boxes: 3.3°, border width: 0.2°; see Figure 2). A yellow left, right, or doubleended (neutral cue) arrow (length: 2°) was presented in the center of the screen. A white target circle (2°) was flashed inside the box for 10ms, either on the side indicated by the arrow (valid cue), the opposite side (invalid cue) or randomly (neutral cue). Distance between fixation and target area border was 7.3°. The fixation sign without cue was presented for a variable duration between 800 and 1000 ms (M = 903.45ms, SD = 58.54ms). Cues were presented for a random duration between 1200 and 1800 ms (M = 1487ms, SD = 174.72ms). Participants were instructed to indicate the appearance of a target via button press (space bar) as quickly as possible. Participants were given a maximum of 1200ms to make a response. Participants completed 120 trials (84 valid (70%), 18 invalid (15%), 18 neutral (15%)).
Data Preparation and Analysis [TOP]
Calculation task [TOP]
Symbolic catch trials were always 16 multiplied or divided by 1. Although responses were jittered, so that ‘16’ was never presented, if participants were paying attention and following the task directions, they should have been able to choose the value closest to 16 most of the time. Some degree of inaccuracy (i.e. < 100%) was expected since response choices were jittered. Therefore, we first eliminated subjects with accuracy less than 50% correct or greater than 3 SD from the group mean. This cutoff eliminated one subject whose accuracy (40%) deviated more than 3 SD from the group mean. In the remaining subjects (n = 17), the average accuracy was between 93% and 70% (M = 85%, SD = 0.36).
Next, because multiplication and division problems were not presented in separate blocks (Katz & Knops, 2014) and subjects might occasionally perform the wrong operation (e.g. multiplication instead of division), data was trimmed to exclude trials where the difference between the log_{10} of the chosen value and the log_{10} of the correct value was more than 3 SD from the subject’s mean. When considering all conditions together, this excluded 40 calculation trials (0.4% of calculation responses, all nonsymbolic). This was less than previous studies using this method (e.g. 1.8%; Knops, Viarouge, & Dehaene, 2009), possibly due to the simultaneous presentation of operands and the use of operation symbols (x, ÷) rather than letters (e.g. ‘A’ for addition & ‘S’ for subtraction; Knops, Thirion, et al., 2009). This likely minimized operation errors. Because a previous study (Katz & Knops, 2014) indicated greater variance for nonsymbolic than symbolic trials, we decided to calculate each subject’s mean separately for symbolic and nonsymbolic notations and to exclude responses beyond mean plus/minus 3SD. This excluded 100 calculation trials (1% of calculation responses, all symbolic). Additionally, in 116 calculation trials (95 symbolic) and 7 catch trials (6 symbolic), no response was made within 6000ms (maximum response duration). The number of these timeout trials ranged from 0 to 37 trials per subject.
Attention task [TOP]
In the Posner task, we first eliminated responses that were faster than 200ms, because these likely reflect premature responses. This eliminated 146 trials (7.3% of responses). The number of responses faster than 200ms per subject ranged from 0 to 37 responses (M = 8.6, SD = 10.4). We then eliminated responses where the RT was more than 3 SD from the subject’s mean. This eliminated an additional 37 trials (2% of valid responses; 14 trials per subject, M = 2.2, SD = 1).
We computed validity effect (RT invalid minus RT valid), benefit (RT neutral minus RT valid) and cost (RT invalid minus RT neutral) as indices for orienting/selection and reorienting, respectively.
Results [TOP]
We first analyzed the distribution of responses to exclude the possibility that participants responded randomly. If participants responded nonrandomly, then range and rank should have a significant effect on response choice (Katz & Knops, 2014). Indeed, we found that rank and range interacted for both multiplication and division in both symbolic and nonsymbolic notation. We report these analyses in Appendix B.
Linear Increase of Response Value With Correct Value [TOP]
Previous studies found that behavior was well described by Weber’s law, suggesting a logarithmic compression of the underlying representation. This also appeared to be true in the present data (see Figure A.2). Transforming data logscale also better meets the prerequisites of ANOVA, stipulating fixed variance (Katz & Knops, 2014). There was a linear increase of response value with correct value in both formats and operations (Table 1).
Table 1
Linear scale

Log scale



t  Slope  95% CI  t  Slope  95% CI  
Multiplication  
Symbolic  123.1**  0.96  0.94, 0.97  169.0**  0.98  0.97, 1.00 
Nonsymbolic  40.7**  1.24  1.18, 1.30  50.4**  1.05  1.01, 1.09 
Division  
Symbolic  129.1**  0.96  0.95, 0.97  148.0**  0.98  0.97, 0.99 
Nonsymbolic  37.0**  0.95  0.90, 1.00  47.9**  0.92  0.89, 0.96 
**Bonferroni corrected for multiple comparisons, p < .013.
Operational Momentum Effect [TOP]
To investigate operational momentum, we entered the response bias, defined as the difference between the log chosen and the log correct values, into an ANOVA comprising the factors notation (symbolic, nonsymbolic) and operation (multiplication, division). There was a significant interaction between operation and notation on response bias (F(1,16) = 6.80, p = .018). Since this makes interpretation of putative main effects difficult we no longer followed up on them. Therefore, simple main effects analysis was used to test the effect of operation, separately for symbolic and nonsymbolic notations. For symbolic problems, operation did not have a significant effect on response bias (F < 1; Figure 3). However, symbolic multiplication (M = 0.007, 95% CI [0.012, 0.002]), but not division (M = 0.006, 95% CI [0.014, 0.001]), was significantly underestimated (i.e. mean response bias < 0; Figure 3).
For nonsymbolic problems, operation had a significant effect on response bias (F(1, 16) = 6.93, p < .05, Bonferroni corrected; Figure 3). The mean response bias was positive for multiplication (M = .023, 95% CI [0.016, 0.062]) and negative for division (M = 0.076, 95% CI [0.127, 0.024]); difference = 0.098, 95% CI [0.019, 0.177]. Whereas we observed a significant underestimation in nonsymbolic division (t(16) = 3.09, p = .007) we found no significant response bias in nonsymbolic multiplication (t(16) = 1.23, p = .238; Figure 3).
Notation did not have a significant effect on response bias for multiplication (F(1,16) = 2.50, p = .134), but did for division (F(1,16) = 7.40, p = .015). The response bias was more pronounced (more negative) for nonsymbolic than symbolic division (t(16) = 2.72, p = .015).
In sum, we replicated the results from Katz and Knops (2014), showing a significant OMeffect for nonsymbolic multiplication and division but no OMeffect for symbolic operations, despite having encouraged approximate calculation by omitting the correct response from the symbolic response alternatives.
Can Attentional Orienting and Reorienting Predict Response Bias? [TOP]
It has been put forward that the OM effect reflects the consequences of an attentioninduced spatial displacement along the mental number line during the process of approximating the outcome of an arithmetic problem (Knops, Thirion, et al., 2009; Knops, Viarouge, & Dehaene, 2009; Knops, Zitzmann, & McCrink, 2013). Here, we tested the straightforward hypothesis of an association between OM and spatial attention as measured in the Posner paradigm.
We first tested whether the relative OM bias, defined as the difference between operationspecific OM bias ((log correct minus log chosen Multiplication) minus (log correct minus log chosen Division)) correlated with the validity effect (valid cue minus invalid cue).
The validity effect was consistently larger than zero (i.e., faster response for valid than invalid) for all participants (M 27.5 ms; t(16) = 5.64, p < .001). To determine whether the validity effect could predict the difference in response bias between multiplication and division, each subject’s validity effect was used as a predictor for the mean logscale relative response bias (multiplication response bias minus division response bias) in a linear regression model. We restricted our analyses to the nonsymbolic notation since symbolic OM bias was not significant. The advantage of a valid cue compared to an invalid cue significantly predicted nonsymbolic relative response bias (F(1,16) = 6.90, p = .019. It accounted for 31.5% of the variability in response bias and had a large effect size ( $\overline{R}$^{2} = .269).
To further examine the attentional mechanisms potentially driving the OM effect in nonsymbolic notation, we separately examined the effect of attentional orienting (benefit of valid cue compared to neutral cue; M = 8.3 ms, t(16) = 2.79, p = .013) and reorienting (cost of invalid cue relative to neutral cue; M = 19.2 ms, t(16) = 4.13, p = .001)) on response bias using linear regression. The adjusted R^{2} value ( $\overline{R}$^{2}) was used to determine effect sizes using the cutoffs: small = .01 or 1%, medium = .1 or 10%, and large = .25 or 25% (VachaHaase & Thompson, 2004). Invalid cue cost significantly predicted relative response bias in nonsymbolic problems; F(1,16) = 13.02, p = .002, $\overline{R}$^{2} = .464. Valid cue benefit could not predict response bias difference in nonsymbolic problems; F(1,16) = 0.23, p = .638, $\overline{R}$^{2} = .044. Results are shown in Figure 4. The correlation between cost and response bias was significantly more negative than the correlation between benefit and response bias (Hotelling’s t(13) = 2.15; Steiger’s Z = 1.87, p < .05 (onetailed)).
All correlations remained byandlarge unchanged after partialing out age as a potential confound (r(validity effect, response bias) = .613, p = .012; r(benefit, response bias) = .088, p = .745; r(cost, response bias) = .704, p = .002).
These results suggest that in nonsymbolic problems, attentional shifts, most likely reorienting but not orienting, largely account for the difference in response bias between multiplication and division.
Discussion [TOP]
In this study we examined two questions. First, we explored whether the previously reported absence of an OM effect in symbolic multiplication and division (Katz & Knops, 2014) may have been due to the presentation of the correct outcome among the response alternatives which may have triggered verbally mediated fact retrieval and hence lowered the impact of visuospatial processes. In the present study, we found no OM effect in symbolic notation even though we encouraged approximate calculation by presenting exclusively incorrect symbolic response alternatives. In contrast, a significant OM effect was observed for nonsymbolic notation, replicating Katz and Knops (2014). Second, we explored the underlying mechanisms of the OM effect by testing the association of the OM effect with visuospatial attention as measured by a Posner paradigm. We found that reorienting attention after the presentation of an invalid cue to the location of the target significantly correlated with the extent to which participants over and underestimated the outcomes of nonsymbolic multiplication and division problems.
The absence of a significant OM in symbolic multiplication and division in the current study is in line with previous results and implies that symbolic multiplication and division strongly rely on verbally mediated fact retrieval which is less prone to cognitive biases such as the OM. While a recent study described how the compression of the mental number line biases arithmetic fact retrieval (Didino, Knops, Vespignani, & Kornpetpanee, 2015), these spatial biases may be too subtle for the current paradigm.
We demonstrated a significant correlation between nonsymbolic OM and measure of reorienting attention after invalid cues in the Posner paradigm. No correlation was observed between OM and orienting attention after valid cueing. This is somewhat unexpected since the attention account of OM holds that attentional shifts propel participants too far along the mental number line. Hence, the benefit from a valid cue should correlate positively with OM. Yet, we did not observe any correlation between OM and cue benefit, a measure of attentional orienting. This might be due to a reduced variability in participants’ performance in our Posner task which had slightly longer SOAs compared to the classical Posner paradigm. Variability may have been particularly reduced for the benefit measures as compared to cost measures, which, all else being equal, lowers correlation coefficients.
Reorienting is not a unitary process but has traditionally been subdivided into disengaging, shifting and reengaging attention to the new location (Posner & Petersen, 1990). In support of disengagement as a separate attentional mechanisms involved in reorienting, recent ERP studies revealed circumscribed posterior components linked with disengagement in the absence of attentional shifts (P4pc, Toffanin, de Jong, & Johnson, 2011), and separate from attentional selection (reversed N2pc, Eimer & Kiss, 2008). This more complex process may be subject to greater amount of variability across participants which, in turn, allows for higher correlation coefficients. Since attentional shifts and engaging are involved in both orienting and reorienting, the absence of a correlation between orienting and OM may as well imply that those processes that are unique to reorienting are at the heart of this association, namely disengaging. Future studies may use EEG to investigate this hypothesis.
Another question that arises from the current findings concerns the direction of the observed correlation. Why is the correlation between OM and reorienting positive, meaning that people with a large OM effect, i.e. larger deviations from the correct outcome exhibit larger costs for invalid cue both compared to neutral cues and valid cues? According to the above theoretical accounts of the OM we can break this question down into four aspects. Why would attentional reorienting correlate with (a) a heuristic according to which multiplication leads to larger outcomes and division to smaller outcomes, (b) flawed decompression, (c) competing spatial biases by the operands, the results or the outcome, or (d) attentional shifts along the mental number line?
According to the first account, no correlation with attentional measures would have been predicted. If any, there would have been a prediction that the expectation of larger or smaller outcomes creates attentional shifts to the left or right which might generate a coarse approximation of the result that can be used to check if a given outcome is plausible or not. However, what renders this hypothesis rather unlikely to account for the given results is that no gradation beyond a coarse “more or less” expectation is predicted.
Similarly, according to the compressiondecompression approach OM results from flawed compressiondecompression mechanisms. No attentional mechanism is involved in this process. According to this account, OM would scale with the size of both operands and results. However, previous research found that OM increases with increasing outcome (Knops, Viarouge, et al., 2009) but no association with operand magnitude was observed. In pointing tasks, OM was strongest when the second operand was zero, clearly speaking against any association between second operand size and OM (Pinhas & Fischer, 2008). Although the compression of the MNL is thought to be logarithmic, it is also subject to interindividual variance. A recent study tested whether the degree of compression of the approximate number system can actually serve to predict OM (Knops et al., 2014). However, the combination of crucial approximate number system parameters such as compression (as measured by the amount of underestimation in nonsymbolic estimation) and precision (as measured by the Weber fraction) in a psychophysical model was not successful in predicting OM. While overall biases in addition and subtraction involving nonsymbolic quantities were well predicted by the interindividual variability of the parameters describing the approximate number system in the model, the operationspecific OM was not. Finally, it is hard to see how this would conceptually relate to the costs of reorienting attention to an invalidly cued target. Further evidence against a heuristicbased account comes from the study of Magethis and colleagues (Marghetis et al., 2014) who found systematic biases in mouse pointer trajectories when participants indicated the correct outcome for addition and subtraction problems. Exploiting the fact that ontask tracking of mouse trajectories provides a timeresolved window on cognition, Margethis and colleagues found that the time course of this spatial deviation sequentially reflected the serial impact of first operand, operator and second operand. The authors conclude that neither a heuristicbased nor a compression account would predict this pattern of results, which is in line with a spatial account of the OM (Marghetis et al., 2014). Together, this suggests that the OM is most likely not fully accounted for by flawed compressiondecompression mechanisms and suggests the origin in parameters outside the ANS.
With respect to the spatial accounts of the OM, a clear prediction comes from the attentional shift hypothesis which predicts a clear association between attentional parameters and OM. The current results partially confirmed this by the correlation between OM and reorienting, providing further evidence for a role of spatial attention during approximate arithmetic and, more specifically, for the idea that the OM results from attentional mechanisms. Larger reorienting costs may reflect highly efficient orienting mechanisms that need to be overruled after invalid cueing. These results are also in line with the finding that symbolic addition problems are solved faster when the second operand is presented on the right compared to leftsided presentation (Mathieu, Gourjon, Couderc, Thevenot, & Prado, 2016). For subtraction an analog advantage for leftsided operands was observed. Crucially, these authors also failed to find a benefit for lateralized operands in multiplication problems, highlighting the different cognitive processes that contribute to multiplication on the one side and addition or subtraction on the other. Recent results corroborate a tight link between the ocular movement system and OM (Klein et al., 2014). Relative to the first fixation, participants subsequently moved their eyes to the right for additions and to the left for subtractions, paralleling previous fMRI results (Knops, Thirion, et al., 2009). This dynamic process of adjusting fixation during the course of arithmetic processing may in part have been specific to the task which required indicating the outcome by pointing to the respective location on a number line. It is also conceivable, however, that approximate calculation is mediated by the dynamic updating within a coordinate transformation system in parietal cortex.
The spatial competition account, in contrast, does not predict any correlation between attentional shifts and OM. According to this account OM is largest when competition between spatial positions of operands and the results on the MNL are minimal. The strong OM in zero problems where only the first operand and the operational sign induce a spatial bias is explained by the absence of a spatial bias induced by zero (Pinhas & Fischer, 2008; Pinhas et al., 2015). Zero is either not represented on the MNL or triggers rulebased procedures. While we cannot test this prediction in the current experiment, we may interpret reorienting in terms of spatial competition between a cued position and the appearance of a target at an uncued position. The spatial competition account would predict that less competition is associated with larger OM. However, we observe the opposite pattern of results in the current study where larger amount of spatial competition is associated with larger OM. Hence, under the premise of interpreting the reorienting effect in terms of spatial competition our results provide evidence against this account. It should be noted, however, that the spatial competition account was initially proposed in the context of addition and subtraction. It is unclear whether it would also hold for multiplication and division where the split between operands and results is much larger.
Limitations [TOP]
As this study provides only correlational evidence, no causal inference can be drawn. Further experimental work is required to elucidate the neurocognitive mechanisms underlying the OM effect. The absence of significant correlations between operational momentum bias and attentional measures may in part be due to the long SOAs in our version of the Posner task. Future studies could increase variability by using shorter SOAs, which might favor finding higher correlations. The fact that we related accuracybased operational momentum bias with speed–based measures of the attentional capacities might further have reduced our statistical power. The limited sample size may raise concerns about (a) stability and reliability of the data and (b) to what extent the observed correlations were due to the increased variability in our sample. We checked whether our results were driven by some outliers by separately excluding all possible combinations of 1, 2 or 3 participants from the sample. We found single participants to have only a minor impact on the correlation pattern. Even when excluding the two participants with the most extreme value pairings, correlations by and large remained significant or marginally significant. To protect against the possibility that age was a confounding variable that drives our results, we recalculated the main findings after partialing out age and found the major findings of the study unchanged. Finally, while one may be concerned about the role of counting in computing the approximate results, we would reason that counting does play a major role in explaining our results. These additional analyses are can be found in Appendix C. Albeit OM and attention appear to be functional related on a conceptual level, future studies could use more comparable parameters for measuring these concepts.
Conclusions [TOP]
To sum up, we failed to observe an OM effect in symbolic multiplication and division. This is in line with previous findings (Katz & Knops, 2014; Mathieu et al., 2016) and suggests that verbally mediated retrieval of arithmetic facts from longterm memory is less prone to spatial biases, yet not immune (Cavdaroglu & Knops, 2016; Didino et al., 2015). In contrast, we found a significant OM effect in nonsymbolic notation, reflecting that participants systematically overestimated results of multiplication problems while underestimating results of division problems. The nonsymbolic OM effect correlated with attentional parameters measured in a Posner paradigm. By differentially analyzing benefit and cost measures we found this correlation to mainly results from the costs of reorienting attention after invalid cues. While the exact mechanisms driving this correlation remain elusive, these results provide further evidence for the attentional shift hypothesis.