Systemizing is defined as the drive to construct systems, to analyze the variables in a system, and to derive the underlying rules that govern the behavior of a system (Baron-Cohen, Richler, Bisarya, Gurunathan, & Wheelwright, Reference Baron-Cohen, Richler, Bisarya, Gurunathan and Wheelwright2003). This allows someone to make predictions about how a system will behave, and to control the system. The empathizing systemizing (E-S) model of sex differences suggests that males on average spontaneously systemize to a greater degree than do females (the reverse being true for empathizing) (Baron-Cohen, Wheelwright, Griffin, Lawson, & Hill, Reference Baron-Cohen, Wheelwright, Griffin, Lawson, Hill and Goswami2002). The Systemizing Quotient (SQ) (Baron-Cohen et al., Reference Baron-Cohen, Richler, Bisarya, Gurunathan and Wheelwright2003) was developed to measure individual differences in the drive to systemize. The aim was to tap into the following types of systems: Technical, natural, abstract, social, organisable, and motoric. The SQ has been translated into fifteen other languages, including Spanish, French, Finnish and Chinese.
The SQ was a 40 item measure with a four choice response structure ranging from definitely agree to definitely disagree. However, Wheelwright et al. (Reference Wheelwright, Baron-Cohen, Goldenfeld, Delaney, Fine, Smith and Wakabayashi2006) argued that the items in this scale were derived “primarily from traditional male domains” (p. 48). In order to guard against circularity – that males might be more likely to score higher on the SQ because the domains are more male-typical – the final 75-item version of the SQ-R (i.e., the Revised version) includes examples of systemizing in varying degrees from everyday life, measuring social and domestic systems, as well as mechanical and abstract systems. Males on average score significantly higher on the SQ-R, relative to typical females (Baron-Cohen et al., Reference Baron-Cohen, Richler, Bisarya, Gurunathan and Wheelwright2003; Wheelwright et al., Reference Wheelwright, Baron-Cohen, Goldenfeld, Delaney, Fine, Smith and Wakabayashi2006). Individuals diagnosed with an autism spectrum condition (ASC) on average score significantly higher than people from the general population. These results have been replicated in a Japanese population (Wakabayashi et al., Reference Wakabayashi, Baron-Cohen, Uchiyama, Yoshida, Kuroda and Wheelwright2007). Child and adolescent versions of the SQ-R have been developed (Auyeung, Allison, Wheelwright, & Baron-Cohen, Reference Auyeung, Allison, Wheelwright and Baron-Cohen2012; Auyeung et al., Reference Auyeung, Wheelwright, Allison, Atkinson, Samarawickrema and Baron-Cohen2009), showing similar patterns of sex differences to those observed in adults, as well as group differences between individuals with and without ASC. SQ scores correlate with increased activation in brain regions associated with increasing and maintaining attention, and do not correlate with IQ (Billington, Baron-Cohen, & Bor, Reference Billington, Baron-Cohen and Bor2008; Ling, Burton, Salt, & Muncer, Reference Ling, Burton, Salt and Muncer2009). Nettle (Reference Nettle2007) also found that systemizing is not explained by existing personality constructs (the Big Five factor model).
To date, few studies have examined the psychometric properties of the SQ. A Principal Components Analysis (PCA) revealed 11 factors (from 40 items) that did not correspond to factors with any obvious psychological significance (Baron-Cohen et al., Reference Baron-Cohen, Richler, Bisarya, Gurunathan and Wheelwright2003). Wakabayashi et al. (Reference Wakabayashi, Baron-Cohen, Wheelwright, Goldenfeld, Delaney, Fine and Weil2006) found 25 of the original 40 items loaded on to one factor and suggested that the scale measures one factor. Ling et al. (Reference Ling, Burton, Salt and Muncer2009) conducted a confirmatory factor analysis (CFA) on 40 items, finding a one factor model was a poor fit. They also examined Wakabayashi et al.’s (Reference Wakabayashi, Baron-Cohen, Wheelwright, Goldenfeld, Delaney, Fine and Weil2006) 25 item unidimensional model, which was also a poor fit. Further analyses led to their suggestion that the SQ is best considered as a four-factor 18-item scale, with subscales measuring topography, technicity, home improvements, and structure, with a hierarchical factor of systemizing. All analyses reported above were conducted on the original version of the SQ (40 items), which may contain items that have a male bias. These results indicate that overall the SQ’s structure is not clear and that confirmatory factor analysis suggests a one dimensional structure is a poor fit to the data.
The purpose of the current study was to re-examine the dimensionality of the SQ-R (75 items) in a larger sample, using CFA, as well as a contrasting statistical approach: Rasch analysis. It was also intended to investigate the possibility of differential item functioning between male and female participants. In particular, to test for bias towards males, or indeed females, that may in appropriately influence scores on systemizing. Although Wheelwright et al. (Reference Wheelwright, Baron-Cohen, Goldenfeld, Delaney, Fine, Smith and Wakabayashi2006) intended to develop a scale without a male item bias; so far this has not been tested.
Rasch analysis is designed to produce unidimensional measures (measuring one ability/personality trait/attitude), which is achieved when the data fit the model. A Rasch scale implies that a sum score can legitimately be used to quantify the trait. Rasch’s (Reference Rasch1960) technique creates scales that fulfill the requirements of additive measurement (Perline, Wright, & Wainer, Reference Perline, Wright and Wainer1979), thus treating ordinal data as interval. A person who has a greater ability than another person should have a greater probability of solving any item of that type in question. The probability of solving an easier item is greater than the probability of solving a harder item. The probabilistic relationship between person ability and item difficulty as a latent trait is modeled. Person ability (percentage of ‘correct’ items) and item difficulty (the proportion of participants who get the item ‘correct’) are located along the same continuum in logits (log odds), transforming data obtained from ordinal scores into interval level measurement. The probability of getting any item correct is produced by the difference between a person’s ability and the item difficulty. If a person’s ability is lower than an item’s difficulty, then the participant is less likely to get this ‘correct’ than if it is higher than the item’s difficulty. This information can be used to compare the actual data collected with expected calculations of item difficulty and person ability. The closer the actual results are to the predicted results, the better fit the data are to the Rasch model.
In the case of the items of the SQ-R we are not concerned with whether an item is correct, but on how characteristic a positive response would be for high and low ‘systemetizers’. A participant with high overall scores on systemizing would be expected to definitely agree with items that are measuring this dimension. The logic and statistical procedures, however, are the same as if the SQ-R was an ability measure.
Rasch analysis will be used to try to develop a one dimensional scale to measure systemizing. It also provides a good method to examine differential item functioning by comparing item responses for male and female participants, so that we can be sure that the scale is not biased towards either sex. Lastly, confirmatory factor analysis will also be used to examine the factor structure of items that fit the Rasch model.
Method
Data source
Data were collected at the websites of the Autism Research Centre (ARC), University of Cambridge. Individuals register as research volunteers and complete online questionnaires and tests. The ARC website Footnote 1 recruits individuals with ASC as well as parents of children with ASC Footnote 2 . Everyone is invited to complete the Systemizing Quotient - Revised (SQ-R). N = 4058 individuals completed the SQ-R online, of which N = 2768 were female and N = 1290 were male. Within this sample, N = 675 individuals had a diagnosis of ASC, N = 1369 were family members of an individual with ASC, and N = 2014 had no diagnosis of ASC. The mean age of the whole sample was 38 years (SD = 12 years). The total sample was randomly divided into two (N = 2029 each), so that the analyses could be independently validated in a new sample. When the whole sample is used this is indicated in the Results.
Material
SQ-R
The SQ-R consists of 75 statements to which participants have to indicate the degree to which they agree or disagree. There are four response options: ‘strongly agree’, ‘slightly agree’, ‘slightly disagree’, ‘strongly disagree’. ‘Definitely agree’ responses score two points and ‘slightly agree’ responses score one point on half the items, and ‘definitely disagree’ responses score two points and ‘slightly disagree’ responses score one point on the other half. The remainder of the response options score 0. See Baron-Cohen et al. (Reference Baron-Cohen, Richler, Bisarya, Gurunathan and Wheelwright2003) and Wheelwright et al. (Reference Wheelwright, Baron-Cohen, Goldenfeld, Delaney, Fine, Smith and Wakabayashi2006) for full details.
Procedure
Rasch analysis
Initial analyses took place on Group A, who were randomly selected from the whole sample (N = 2029), and was then replicated in Group B (N = 2029). The Rating Scale (Andersen, Reference Andersen1977) routine in WINSTEPS (Linacre, Reference Linacre2006) was conducted. The WINSTEPS reliability estimate was executed to provide an estimate of cohesion of the items (person and item reliability estimates). Point-biserial correlations between items scores and total score were inspected. In Rasch analysis point-biserial correlation is used as a first check that the scoring makes sense as all items should be positively correlated (Linacre, Reference Linacre2006). More importantly, the fit of each of the items to the model can also be investigated. There are two fit measures, Infit and Outfit, which both reflect the ratio of the observed variance to what we would expect from the Rasch model. Infit, however, is weighted to reduce the effect of outliers or extreme responses. Item and person misfit and item Infit and Outfit statistics were examined. PROX estimation was used initially to converge the data with the Rasch model. Unconditional maximum likelihood estimation (UCON) was subsequently used.
A principal components analysis (PCA) was performed on the residuals after the Rasch measurement factor was removed. This is conducted to check for the possibility of other large components that explain the data, other than the Rasch dimension (Linacre, Reference Linacre1998). If there is such a component then the data do not fit a unidimensional model. The components/factors were examined to check for unidimensionality and to assess whether a single latent trait explains the majority of the variance in the data (Mavranezouli, Brazier, Young, & Barkham, Reference Mavranezouli, Brazier, Young and Barkham2011). The ratio of variance explained by the Rasch measurement factor to that explained by the residual factors was also analyzed (Wright & Stone, Reference Wright and Stone2004).
Differential item functioning can be investigated with Rasch analysis by comparing item difficulty measures (in this case likelihood of agreement) from each sex. Items should be invariant across groups. If the ‘item difficulty’ estimates between the two genders shift significantly, then this suggests that the two genders are responding differently to the items and the scale overall is not measuring the same thing in both genders. In order to examine differential item functioning the Rasch analyses were subsequently carried out separately on the male sample and the female sample.
Confirmatory Factor Analysis
CFA was conducted on the full 75 item version of the SQ-R (on Group A) using Amos (Arbuckle, Reference Arbuckle2006) with maximum likelihood estimation. First a one factor solution was specified. The chi square (χ2) value and degrees of freedom, the Comparative Fit Index (CFI) and the root mean square error of approximation (RMSEA) were examined. Browne and Cudek’s (Reference Browne, Cudeck, Bollen and Long1993) criterion for good fit was used, suggesting an RMSEA under .08 represents reasonable fit, and below .05 representing very good fit (Steiger, Reference Steiger1989). Higher values of the CFI are considered better with values over .9 considered acceptable.
The data were then also examined using a CFA parceling approach. A parceling approach was adopted, to reduce the number of items, as fit statistics are affected by the number of items (Kenny & McCoach, Reference Kenny and McCoach2003), and scales with large numbers of items generally have very poor fit (Cook, Kallen, & Amtmann, Reference Cook, Kallen and Amtmann2009; Floyd & Widaman, Reference Floyd and Widaman1995). This is because in confirmatory factor analysis the unique variances of items may be correlated because the items share a specific feature, and this becomes more likely with more items. This is particularly true also when many items are constructed from similar domains, as is the case with the SQ-R. Parcels of items are also more reliable and are also more likely to have linear relations with each other and with the proposed latent factors (Comrey, Reference Comrey1988). Yang, Nay, and Hoyle (Reference Yang, Nay and Hoyle2010) argue that parceling has been identified as a desirable approach when there are more than 12 items in a scale (Marsh, Hau, Balla, & Grayson, Reference Marsh, Hau, Balla and Grayson1998) and when the items reflect a unidimensional construct (Hall, Snell, & Foust, Reference Hall, Snell and Foust1999).
Results
Rasch analysis - Comparison of two sub-samples
Three UCON iterations brought the convergence to about 20% of the starting values (Group A) with four further iterations required to achieve full convergence. This small number of iterations required indicated that the data was a relatively good fit to the Rasch model, particularly given the large number of both items and participants. Item reliability was high at .99, which indicated that the item difficulties (strength of indication of systemizing) were very likely to replicate if the items were given to another sample. The person reliability was also high at .92, which indicated that the ordering of participants in terms of their systemizing was likely to replicate if they were given another parallel set of items. Both of these figures suggested that there was a viable Rasch dimension in the data.
For Group B, item reliability was 0.99 and person reliability was 0.93, following the same pattern of UCON iterations, indicating the SQ-R has exceptionally high item reliability. The items coalesced around a single cohesive variable. The high reliability for persons also indicated that the sample participants responded to the SQ-R items in a cohesive manner. Item calibrations for both samples correlated at r(75) = .99, confirming the consistency of item calibrations across the two samples.
The PCA of the residuals following extraction of the Rasch measurement factor on Group A and Group B produced similar outcomes in each group. Items loaded onto the same components for both groups, and the loadings were of similar strength. The correlation between the two sets of item loadings from group A and group B was r = .87. This indicated that the two randomly derived samples were producing similar results and therefore were combined for further analysis.
Rasch analysis - Total sample statistics
Using the whole sample, item reliability was .99 and person reliability was .92, which suggested that the data were fitting the Rasch model. The point-biserial correlation were generally in the right direction with only item 3 with a negative correlation. More importantly both the Infit and Outfit statistics were indicative of a good fit of all of the items to the Rasch model. A fit value of 1 indicates perfect fit and fit values of over 1.5 are taken to indicate poor fit (Linacre, Reference Linacre1998). None of the items showed poor fit by that criterion, see Table 1.
Table 1. SQ Rasch Item Statistics (75 items)
The PCA of the residuals can also be used to assess the unidimensionality of the SQ-R. The eigenvalue of the residual component/factor is quite high at 5.2 and exceeds Linacre’s (Reference Linacre1998) criterion for the presence of another factor. The factor sensitivity ratio (Wright & Stone, Reference Wright and Stone2004) can be used to evaluate the importance of residual factors after the Rasch measurement factor has been extracted. This is calculated by dividing the residual variance by the Rasch measurement variance. The ratio of the residual factor variance (5.2) to the Rasch measurement variance (21.6) was 0.24. This suggests that about 24% of the measure is affected by unexplained relationships between the items, which is large enough to require explanation. To put it another way about 75% of the variance in the measure can be explained by the Rasch measurement factor, but the other factor is fairly large.
The factor loadings were examined In order to try to define the other factor, besides systemizing, that might be influencing the result. An examination of the items that load positively and negatively at over .3 onto the residual first factor did not suggest that this factor had any psychological meaning, but was more likely to be method variance. All twelve of the items with strong positive loadings on the residual factor were scored in the direction of agreement, and nine of the ten items with strong negative loadings were scored in the direction of disagreement. This strongly suggests that although the SQ-R is measuring one dimension, this is influenced by the direction of scoring of the items. A very similar pattern of results was found by Allison, Baron-Cohen, Wheelwright, Stone, & Muncer (Reference Allison, Baron-Cohen, Wheelwright, Stone and Muncer2011) in their analysis of the original Empathy Quotient scale. On that occasion they pointed out that some researchers had mistaken method variance for something more psychological.
The scores on items scored in the direction of agreement and disagreement were significantly related r(4058) = .68, p < .005, and both types of items showed large sex differences. There was, however, a significant interaction between gender and type of item F(1, 4056) = 27.94, p < .005. The difference between sexes was larger for items scored in the direction of agreement (d = .53), than for disagreement (d = .47).
Differential item functioning
Item and person reliabilities for both males and females were .99 and .92 respectively, which suggests that the scale overall works well for both sexes. One method of comparing item functioning is to conduct a t test comparison of the item measure scores from both sexes. Such a comparison has been shown to be oversensitive with reasonably large samples, so we followed Tristan’s (Reference Tristan2006) procedure for adjusting for sample size. On this basis five items showed differential item functioning. All of these had a very large difference in item ‘difficulty’ (0.8+ logits) between males and females. These were items 20, 32, 43, 49, and 55 (see Table 2 for the full details of SQ-R items).
Table 2. The SQ-R 75 items with indication of Agreement or Disagreement scoring procedure
For these items the gender difference was more than one would expect to be caused by differences in systemizing. Items 20 and 55 are both related to shopping and females scored more, and males less than would be expected. It is possible that these items are affected by the differences in likelihood of shopping by males and females. Item 49 was “I do not tend to remember people’s birthdays (in terms of which day and month it falls)”, where males scored much lower than expected. It is possible that males are responding mostly to the first part of the item, that is they do not remember birthdays, and therefore, the second part which might indicate systemizing is largely ignored. Item 32 (“I am fascinated by how machines work”) and item 43 (“If there was a problem with electrical wiring in my home, I’d be able to fix it myself”) have particularly low scores from females.
It is fair to say that at least four of these items are poorly constructed. Responses to the shopping and birthday items are less indicative of differences in SQ and more related to other sex differences. The item about electricity asks not just about interest but also about a specific ability, which is likely to be more prevalent in males. Given the relatively small number of items that show severe differential item functioning, however, it is probably not necessary to omit these items from the scale. For the current data removal of these items reduced the effect size of gender on SQ total from d = .57 to d = .56. Participants with a diagnosis of Asperger’s syndrome showed a similar pattern but with larger effect sizes; for the full 75 items the effect size was d = .69 and for the reduced version d = .70.
So far we have taken a relatively conservative view of differential item functioning by selecting items based on the modified t test procedure. It is possible to take a stricter definition and this is often done when it is important to rule out bias. For example, Scheuneman and Subhiyah (Reference Scheuneman and Subhiyah1998) used an item measure difference of greater than 0.5 logits to detect bias in a medical certification test given by the National Board of Medical Examiners. For the present data, we also took a stricter criterion of a logit difference of greater than 0.4, as suggesting differential item functioning between males and females. The number of items which were now omitted is 31. There are 44 items still remaining, and 23 of these are scored in the direction of disagreement. These 44 items are indicated by an underlined item number in Table 2. There is, however, still a large and significant sex difference in the expected direction t(4056) = 16.97, p < .005, with a slightly reduced effect size d = .53. There is also a large and significant difference for those with a diagnosis of Asperger’s syndrome t(4056) = 22.83, p < .005, with a larger effect size of d = .71. It is clear that there is a substantial difference between males and females on systemizing, and also between Asperger’s sufferers and non-sufferers, that cannot be ascribed to bias in the items of the Systemizing Quotient.
Confirmatory Factor Analysis
It was useful to examine the data from the two created samples separately as if modifications of the proposed structure were required then these could be validated on the second sample. Fit statistics from AMOS (Arbuckle, Reference Arbuckle2006) for Group A are presented in Table 3. The one factor solution for all 75 items was not a good fit according to the χ Footnote 2 value or the CFI, despite a reasonable RMSEA value. This is not surprising, given the problems in using CFA at item level on large tests. The items with the most problematic loadings onto a single factor were 3 (–.09) and 49 (.05). Fifty-eight of the items had loadings onto the latent factor that were greater than .3.
Table 3. CFA of the SQ – Group A
There are a number of different possible approaches to parceling (Landis, Beal, & Tesulk, Reference Landis, Beal and Tesulk2000). These include parceling by content in which items are parceled according to their item content, parceling based on an exploratory factor analysis in which items are parceled according to factor loadings, and random parceling. As the Rasch analysis suggested that a one factor solution was possible, a random parceling approach was taken. In this approach it is assumed that each of the items is an indicator of the factor and therefore any combination of items should produce parcels which will be a good fit to a one factor solution. It was decided to create 5 parcels, which insured an equivalent number of items in each parcel, and the number of parcels would not be too large. When the 75 items were randomly parceled into five groups, the CFI suggested a good fit (see Table 3). The correlations between parcels ranged between.71 to .78. The RMSEA value of .11, however, was still higher than would be expected for a well-fitting model. The modification indices revealed that the model could be improved by allowing correlated error between parcels 2 and 4. This resulted in the CFI increasing to .99, with the RMSEA now at .05, indicating excellent model fit to the data.
This model with correlated error between the same parcels was then tested on the data from Sample B. The pattern of results was very similar in Group B. For the 75 item SQ-R with parceling, the CFI was .99 and the RMSEA was .044, when correlated error was allowed between parcels 2 and 4. When the factor loadings, covariance of the error term between parcels 2 and 4, and the variance of the latent variable were constrained to be the same for both groups of participants, the fit of the constrained model was not significantly different to the model in which these parameters could be independently estimated (Δχ Footnote 2 (6) = 8.93, p > .05). This indicated that the proposed structure of the items and their loadings are relatively consistent across samples.
Aluja and Blanch (Reference Aluja and Blanch2004) have argued that a Single Factor item parceling method works well. In this procedure parcels are formed based on the factor loadings when a single factor is extracted. Parcels are formed by selecting items with high and low loadings onto the factor alternately. This parceling method was used to select five different parcels and the fit also suggested that SQ could be seen as one factor (χ2(4) = 41.2; CFI =.99; RMSEA =.048).
Parceling was also tested on the 44 items without differential item functioning between the sexes. These items were randomly split into four parcels to maintain equivalent numbers between parcels, and the fit statistics were acceptable (χ Footnote 2 (2) = 30.09, p < .05; CFI =.997; RMSEA = .06). The factor loadings for all four parcels were also greater than .82. This pattern of results increases support for the proposition that the SQ is measuring a single dimension and should be treated as such.
Discussion
This study examined the psychometric properties of the SQ-R, and explored in particular whether the latent construct of systemizing is unidimensional. Results from the Rasch modeling and CFA with parcels indicated that the SQ-R measures a single dimension of systemizing, and it is therefore appropriate to use a summed SQ-R score to describe the extent to which an individual possesses a drive to systemize.
From the Rasch analysis, high item calibration reliability values were found for both male and female participants. Some items, however, showed differential item functioning by sex. When these items were omitted, however, the overall sex difference was very similar suggesting that differences cannot be ascribed to the biased items and that any differences are a consequence of sex differences in systemizing. It is possible, therefore, either to use the full 75 item version or to use the shorter 44 item version which removes items that show differential item functioning. The SQ-R in either the 75 or 44 item version, also shows differences in the expected direction between those with and without ASC.
The PCA on the residuals remaining from the Rasch analysis suggested that participants are affected by the direction of the items, as the first factor located Agree and Disagree items on opposite ends of this factor. This appeared to be the main defining element of this factor. It should be noted that such measurement factors are quite common. Allison et al. (Reference Allison, Baron-Cohen, Wheelwright, Stone and Muncer2011) previously argued that measurement factors on the Empathy Quotient scale had been misinterpreted as being meaningful psychological factors.
This is not to say that it would be impossible to find subfactors that might be argued to exist. For example, an examination of the modification indices deriving from the CFA of the 75 item structure, suggested things could be improved by allowing correlated error between some items. For example, correlated error between item 20 Whenever I run out of something at home I always add it to a shopping list and item 72 When I have a lot of shopping to do I like to plan what shops I am going to visit and in what order, would improve the fit. However, this seems to be based merely on the subject of the item, ‘shopping’ and not a psychological factor of any interest. The pattern of modifications reveals a number of these item content clusters. We would argue that they have no psychological relevance and are in fact examples of bloated specifics, to which both Exploratory and Confirmatory Factor Analysis are particularly susceptible. We would argue that these are spurious results based on overlap of item content, and agree with Nunnally and Bernstein (Reference Nunnally and Bernstein1994, p.316) that
“…spurious results may lead to inappropriate criticism of sound scales or, what is basically the same thing, lead an investigator to falsely believe that the scale that he or she has developed is appropriately multidimensional when in fact it is not.”
A parceling approach was an efficient way of reducing the number of items on the SQ-R, of overcoming the problem of ordinal data and producing data that is closer to a normal distribution. It was also less likely to be affected by unimportant but specific aspects of item content. CFA using parceling of the items in both the 75 item and Rasch-produced 44 item versions with no differential item functioning resulted in strong fit statistics, indicating that both versions of the scale are best viewed as a unidimensional measure of systemizing. It should also be remembered that these results were independent of method of parceling, as both random and the single factor method (Aluja & Blanch, Reference Aluja and Blanch2004) of parceling produced acceptable fit statistics.
As Nettle (Reference Nettle2007) highlights systemizing appears to emerge as a cognitive style, whereby people are drawn to understanding causal relationships in non-social domains, which can explain the greater male interest and performance in subjects such as science and mathematics (Geary, Saults, Liu, & Hoard, Reference Geary, Saults, Liu and Hoard2000), and the extreme end of the typical male profile in ASC (Baron-Cohen, Reference Baron-Cohen2002). No data are available that assess whether an intact or increased drive to systemize holds true right across the whole autism spectrum. However, repetitive behaviors (including what is described as ‘stimming’ (Wing, Reference Wing1997) as well as collecting obsessions, lining things up, and being drawn to repeating patterns) may be the hallmark of strong systemizing in individuals with ASC and severe learning disability (Baron-Cohen, Reference Baron-Cohen, Zelazo, Chandler and Crone2009).
The current research using two different but complementary perspectives and statistical methods (Rasch modeling and CFA), both highlight that the construct of systemizing is unidimensional and can be measured using the SQ-R. Thus, using a total score on the SQ-R adequately represents and quantifies an individual’s drive to systemize. These data support Baron-Cohen’s (Reference Baron-Cohen2002) construct of systemizing as a drive. There are items in the SQ-R that enquire about preferences, understanding, and to some extent abilities, but these all appear to relate to one single underlying construct of a drive to construct systems, to predict the behavior of a system, and to control it. This study does not address whether a relationship exists between systemizing and intelligence. Previous research suggests this is not the case (with intelligence measured by the Baddeley 3 minute reasoning test) (Ling et al., Reference Ling, Burton, Salt and Muncer2009). Further research is required with more standardized measures of IQ, to properly address this question. Systemizing (in conjunction with empathy measures) is a better predictor than sex of whether a student opts to study science vs. humanities (Billington et al., Reference Billington, Baron-Cohen and Bor2008), suggesting that cognitive style (independent of sex) underlies choice of degree. Whether the drive to systemize is directly related to an ability to systemize, remains untested. Do people with a strong drive towards systemizing always demonstrate high levels of performance on measures that are designed to directly measure and quantify individual differences in systemizing? Can these individuals be separated from individuals with a high drive to systemize, but low levels of performance on systemizing measures? Robust performance measures of systemizing are required to address these questions. Further, performance measures of systemizing that are applicable to the whole population, including people with autism and intellectual disability should be developed. In summary, two different statistical approaches support the idea that the SQ is an appropriate measure of the construct of systemizing which can be measured along a single dimension.
