Sex differences in repetitive stereotyped behaviors in autism: Implications for genetic liability

Abstract

The implications of the well known sex differences in the prevalence of autism spectrum disorder (ASD) are not well understood. The aim of this paper was to investigate whether these differences might be associated with differences in genetic liability. Individuals with ASD (970 families, 2,028 individuals) were recruited as part of the Autism Genome Project (AGP). The families were differentiated into families containing a female (either female–female or male–female) and those with only males. If the sex with the lower prevalence is associated with a greater genetic liability necessary to cross sex-specific thresholds, the males from female containing families should be more severely affected than males from male only families. Affected subjects from the different types of families with ASD were sampled and compared on the social reciprocity and repetitive behavior scores from the Autism Diagnostic Interview-Revised (ADI-R). In general, females had lower repetitive behavior scores than males. More importantly, males from female containing families had higher repetitive behavior scores than males from male–male families. No such differences were apparent on the social reciprocity scores. These results support the hypothesis of a multiple threshold model of genetic liability of ASD with females having a higher liability for affectation status, at least on the repetitive behavior dimension of the disorder. These data also support the dissociation of the different phenotypic dimensions of ASD in terms of its genetic architecture. The implications of these results for linkage and association studies are discussed. © 2011 Wiley Periodicals, Inc.

INTRODUCTION

Autistic disorder is the most severe of a group of disorders, collectively known as the pervasive developmental disorders (PDD) [APA, 1994] or the autism spectrum disorders (ASD). Autism usually presents in childhood with impaired reciprocal social interaction and communication and a pattern of repetitive behavior and/or circumscribed interests. The most recent estimate of the prevalence of autism is at least 30 per 10,000 [Fombonne, 2009] and the other forms of ASD may be as common, leading to an overall prevalence of 1 per 150. The most consistent finding in the epidemiology of autism is that the ratio of boys to girls is about four to one and there appears to be differences in the frequency of severe intellectual disability (ID) by sex [Lord and Schopler, 1985; Volkmar et al., 1993]; that is, an over-representation of girls with IQ's in the severe to profound intellectual disability range where the sex ratio approaches unity. The ratio is apparently greater than 4:1 among children with IQ's above 70 [Lord and Schopler, 1985; Volkmar et al., 1993; McLennan et al., 1993]. This, however, only appears to be true in families with a single affected child; in multiple incidence families the 4 to 1 sex ratio is constant across IQ strata [Banach et al., 2009]. It is also important to note that studies that report sex differences in IQ tend to be older and rely on DSM III rather than DSM IV diagnoses. In addition, the pattern of sex differences in more specific cognitive abilities appears more complex than that observed using a composite measure of IQ [Carter et al., 2007].

There are also some data showing a sex difference in autistic symptoms, (particularly as measured by the Autism Diagnostic Interview; the ADI) but the findings are inconsistent. For example, both Lord et al. [1982] and McLennan et al. [1993] found that boys showed more repetitive stereotyped behaviors than girls. On the other hand, neither Carter et al. [2007] nor Holtmann et al. [2007] found differences in autistic symptoms, particularly repetitive stereotyped behavior, by sex. However, these studies were of relatively small sample sizes and may have lacked power to detect small but important differences. In the former study [Carter et al., 2007], the lack of differences may also be due in part to the young age of the sample since repetitive stereotyped behaviors appear to emerge somewhat later in development. It is also relevant to point out that sex differences reported on the ADI-R may be a function of parent report being influenced by gender stereotypes. However, in the most recent report using the Autism Diagnostic Observation Schedule (ADOS), with a young sample of ASD children 1.5 to 4 years of age, boys were found to show higher restricted, repetitive behavior scores than girls [Hartley and Sikora, 2009]. Thus, no definite answer emerges on sex differences in autistic symptomatology and no convincing explanation of variation in some aspects of the phenotype (such as IQ) but not necessarily others (i.e., autistic symptoms) exists.

One possibility is that these sex differences in prevalence and possibly symptoms are related to genetic factors. The genetic epidemiologic data in ASD cannot be explained by any single gene mechanism so genetic heterogeneity, reduced penetrance, epigenetic factors, and the involvement of multiple genes, as well as gene–gene and gene–environment interactions need to be considered. Of these, most evidence exists for the involvement of multiple genes. The differences in concordance rates between MZ and DZ twins and the fall off in risk from first to second to third degree relatives is consistent with the operation of multiple genes [Szatmari et al., 1998]. The published genome wide linkage [for a review see Losh et al., 2008] and association scans [Ma et al., 2009; Wang et al., 2009; Weiss and Arking, 2009] have not detected major gene effects but some genes of smaller effect are now being identified [Abrahams and Geschwind, 2008]. In addition, relatively rare de novo copy number variants (CNV's) represent potentially important risk factors for a small number of cases but the proportion of cases that can be attributed to this mechanism remains to be determined [Cook and Scherer, 2008].

There are two possible interpretations of these observed sex differences in terms of genetic factors. First, a variation of the polygenic multifactorial (PMF) model seems to fit some of these data quite well (though not without qualification, see below). This model hypothesizes that multiple genes combine and interact in either a multiplicative or additive way to cross a threshold of liability, which results in the expression of the disorder [Tsai and Beisler, 1983; Lord and Schopler, 1985; Volkmar et al., 1993]. The PMF model can incorporate sex-specific thresholds across a normally distributed continuum of liability for males and females. Females are hypothesized to have a higher threshold and must therefore inherit a “larger genetic load” [or a second “hit” see Mefford and Eichler, 2009] in order to overcome this higher threshold. The higher threshold leads to a lower prevalence. However, the difference in IQ between males and females with autism could not be explained fully as a result of this larger genetic “dose.” In other psychiatric disorders with a sex difference in prevalence the less frequently affected sex is also less severely affected. For example, schizophrenia also tends to be less common in females than males and females are less severely affected than males [Roy et al., 2001]. The data on other non-psychiatric diseases such as asthma, lupus, or ankylosing spondylitis show the opposite. For example, asthma is more common in males but more likely to persist in females [Sears et al., 2003]. In lupus, females are more frequently affected, whilst males are more severely affected, and in ankylosing spondylitis males are more frequently affected whilst females tend to have a more severe form of the disease [Ward and Studenski, 1990; Wonuk et al., 2007; Al-Mayouf and Al-Sonbul, 2008]. In sum, sex differences in phenotypic expression may or may not be the result of genetic factors; it appears to depend on the disorder under investigation and on the factors that affect variation in expression which may be specific to that disease.

A second, related, possibility is that of genetic heterogeneity; autism in females may arise from a different set of genetic factors than autism in males. The penetrance of those genes may be less in females than in males and may also lead to greater severity. For example, autism in males could be due in part to one or more loci on the X chromosome acting in a recessive fashion and be associated with a milder phenotype, whereas females with autism could arise as a consequence of a less common autosomal locus or some type of CNV and be associated with a more severe phenotype than one arising from the loci on the X chromosome [Szatmari and Jones, 1991; Zhao et al., 2007]. Some of those affected females will have an affected male relative simply due to the fact that this sub-type represents an autosomal disorder. Thus under this model, males with ASD will be more genetically heterogeneous than females.

Ottman [1987] has proposed a test of the sex specific threshold and the heterogeneity models (they cannot really be distinguished at the behavioral level). This involves a comparison of relatives of male and female index cases and their affected siblings. Under this model as applied to ASD, females with the disorder have a higher genetic liability than most males with the disorder, simply as a consequence of having to cross a higher threshold. Thus, in general, the male relatives of female cases would also inherit a “larger” genetic liability than the male relatives of male index cases. As a consequence, the former group should be more frequently affected than the latter. There are few data on this issue. Tsai et al. [1981] reported that 3% of the relatives of female cases had autism compared to 1% of relatives of boys with autism (for ASD, it was 8% for the relatives of female cases and 3% for the relatives of male cases). Ritvo et al. [1989], in a population-based study, reported that the risk of autism in siblings of girls with autism was 14.5% compared to 7% among siblings of boys. Neither of the differences in these studies reached statistical significance but this might be due to low power either because of the small effect sizes observed, modest sample sizes (N's of 102 and 207, respectively), or both. The most recent report on this issue (N = 417) found no difference in the rates of ASD among male relatives of female cases compared to the rate in male relatives of male cases [Goin-Kochel et al., 2007].

It is also possible to look at quantitative traits that reflect sub-phenotypes of ASD [Szatmari et al., 2007a]. If variation in separate quantitative traits is related to different components of genetic liability, then the relatives of females who have autism should also have a more severe form of the sub-phenotype than the relatives of male cases. The important assumption here is that severity of expression is determined by the same familial (presumably genetic) factors that function as susceptibility or risk factors. The next question is whether the differences in severity due to sex in autism, or one of its sub-phenotypes, are also familial; if so, are they due to genetic heterogeneity, to some variation of the PMF model or are they perhaps due to epigenetic mechanisms that may be sex-specific [Augur et al., 2010]?

This purpose of this study is to test whether the sex differences in severity of quantitative traits seen in autism are familial; are they, in fact, related to differences in genetic liability (recognizing that this is a descriptive level of analysis rather than a test of a causal biological mechanism)? The Autism Genome Project (AGP) data set provides an excellent opportunity to test this since a large sample of multiple incidence families are available which can be divided into those that contain a female affected individual and those in which only males are affected with ASD. Our hypotheses are that, in general, females will have lower scores on ASD behaviors (i.e., have fewer behaviors) than males but that their affected male relatives will have higher scores (have more behaviors) than males from male only families.

METHOD

Sampling

The original 1,397 multiple incidence families were collected from the ten AGP sites in North America and Europe. A detailed description of the AGP data can be found in the first AGP paper [Szatmari et al., 2007b]. Families for the current analysis were included if they had at least two individuals older than 24 months of age and diagnosed with ASD on the basis of a best-estimate clinical diagnosis and the ADI-R. Based on the recommendations of Risi et al. [2006], subjects were regarded as having ASD if they were (1) at/above the ADI-R autism cut-off on the social, communication and repetitive behavior domains; or (2) one point below the ADI-R autism cut-off on both the social and communication domains; or (3) at/above the autism cut-off on the social domain but one or two points below the cut-off on the communication domain; or (4) at/above the autism cut-off on the communication domain but one or two points below the cut-off on the social domain. This sample was most similar to the “ASD all” family sample in our previous report [Liu et al., 2008]. To maximize sample size, we did not require that an ADOS be completed on each subject but all affected individuals were assessed clinically and received a best estimate diagnosis of ASD (of the 2,028 subjects eligible for analysis, 573 did not have an ADOS). All subjects underwent medical evaluations to rule out known neurological and chromosomal disorder. In addition to the above sample exclusion criteria, to reduce genetic heterogeneity caused by ethnicity differences, only families which were inferred to be of Caucasian origin by the computer program “smartpca” in the Eigensoft package (v1.0) were included in the analyses [APA, 1994]. Details of this analysis of population stratification can be found in the paper by Liu et al. [2008]. Of the 1,397 families, 1,182 were of Caucasian origin (215 non-Caucasian were excluded). Of those Caucasian families, 978 had at least two related family members with ASD; 8 families were excluded due to doubtful data and diagnostic criteria. In the end, a total of 970 Caucasian families (2,028 individuals) with at least two related individuals with ASD were included in the analyses [for more information on sampling process see Supplement 1 in Liu et al., 2008]. There were 77 of the 970 families with more than 2 sibs. Of these, there were two families with five sibs, seven with four sibs, and 68 with three sibs.

INSTRUMENTS

Autism Diagnostic Interview-Revised (ADI-R) [Lord et al., 1994]

This is an investigator based interview administered by a trained interviewer to the primary caregiver(s) of the child. It is designed to obtain detailed descriptions of behavior necessary for the diagnosis of autism spectrum disorder (PDD), especially autism. The interview, which is divided into six sections, focuses on the key diagnostic features described in the International Classification of Diseases, 10th Edition [WHO, 1992] and The Diagnostic and Statistical Manual, 4th Edition: general orientating to obtain background information about the subject; early developmental history; communication and language; social development and play; unusual interests and behaviors. An algorithm has been constructed using items from the ADI-R to operationalize the diagnostic criteria for autistic disorder and ASD as outlined in ICD-10 and DSM-IV. Total scores are calculated for social interaction (SOC), and for repetitive behaviors (BEH), with higher scores indicating greater number and severity of ASD behaviors. We did not consider the communication items since there are difficulties in interpretation of communication scores in verbal and non-verbal individuals with ASD.

IQ Measures

The different AGP sites used a variety of different IQ tests depending on the age of the child, their verbal status, whether the estimate was obtained from medical records or by the research team, and local preference. As a result, it was not appropriate to simply combine all the different tests and obtain a single quantitative measure. Instead, each site reviewed the IQ data available and provided a “best-estimate” based on the judgment of a clinical psychologist that was part of the research team indicating whether the subject's IQ was at least 70 or 69 and below. This stratification provided a proxy measure of “higher” and “lower” functioning ASD, respectively.

ANALYSIS

The analysis was carried out in several stages. First, families were divided into female containing (FC) families if they contained at least one affected female with ASD and at least one affected male, female only (FO) families if the only affected subjects in that family were female, and male only (MO) families if all affected individuals in that family were male. From these three types of families there could be four types of affected individuals; females from FC families, males from FC families, females from FO families, and males from MO families.

A mixed linear model with family as a random effect was applied with BEH or SOC scores as a dependent variable and sex of the subject (in two categories—male vs. female); or sex and type of family (the three categories as mentioned above) as an independent variable. This method takes account of multiple affected relatives per family. The BEH and SOC scores were normally distributed in this sample. The model was tested adjusting for the effects of three covariates; AGP site, age of ADI-R completion and verbal/non-verbal status. Verbal status was based on the ADI-R (item number 30). There was no effect of IQ (<70 vs. ≥70) on BEH scores (P = 0.70) so this was not used as a covariate. For SOC, after adjustments for verbal/non-verbal status, AGP site, and age at ADI assessment, IQ accounted for <1% of its total variance so it too was not used as a covariate. Among the 2,028 affected subjects, there was no association between IQ and gender (chi square = 0.19, P = 0.67) nor between IQ and verbal/non-verbal status (chi square = 0.097, P = 0.75) in this sample.

Due to non-normality, Box-Cox transformation was applied to age of ADI-R completion. The least squares means for sex and type of family categories were reported. Post-hoc contrasts were conducted for items within the BEH domain to identify on which items significant differences existed between the groups. SAS was used for the analyses (SAS v 9.1, SAS Institute, Cary, NC).

RESULTS

Of the 970 families selected for the analyses, there were 631 male-only families, 51 female-only families and 288 female-containing families. Table I provides the descriptive statistics for the individuals from all the ASD families and the three subsets of ASD families. All of the three covariates, AGP site, age of ADI-R completion and verbal/non-verbal status, were significantly associated with both SOC and BEH in the mixed linear model (all P < 0.0001). Details of the covariate effects are in the supplemental Table. Variations in both SOC and BEH scores were associated with site and age of ADI-R completion. However, the relationships of SOC and BEH with verbal/non-verbal status were different: the children with ASD who were verbal had significantly higher BEH values than those who were non-verbal. On the other hand, verbal children with ASDs had significantly lower SOC values than those who were non-verbal.

Table I. Descriptive Statistics for the Individuals From All the ASD Families and From the Three Subsets of ASD Families; N's and (%)

	All ASD	Male-only	Female-only	Female-containing
Sex
Male	1,622 (80)	1,302 (100)	0 (0)	320 (52)
Female	406 (20)	0 (0)	105 (100)	301 (48)
Verbal/non-verbal status
Verbal	1,342 (66)	868 (67)	66 (63)	408 (66)
Non-verbal	686 (34)	434 (33)	39 (37)	213 (34)
Best estimate IQ
≥70	920 (66)	608 (66)	46 (65)	266 (66)
<70	478 (34)	314 (34)	25 (35)	139 (34)
Onset of first words
≤24 months	910 (47)	588 (47)	40 (40)	282 (47)
>24 months	1,039 (53)	656 (53)	60 (60)	323 (53)
Onset of first phrases
≤36 months	647 (34)	421 (34)	26 (27)	200 (34)
>36 months	1,261 (66)	803 (66)	70 (73)	388 (66)
AGP site
AGRE	578 (28.5)	345 (26.5)	30 (28.6)	203 (32.7)
VANDERBILT	95 (4.7)	58 (4.5)	4 (3.8)	33 (5.3)
IMGSAC	446 (22.0)	289 (22.2)	25 (23.8)	132 (21.3)
DUKE	99 (4.9)	54 (4.1)	6 (5.7)	39 (6.3)
CANAGEN	183 (9.0)	125 (9.6)	10 (9.5)	48 (7.7)
INSERM	74 (3.7)	41 (3.2)	5 (4.8)	28 (4.5)
STANFORD	164 (8.1)	117 (9.0)	9 (8.6)	38 (6.1)
CPEA	276 (13.6)	188 (14.4)	10 (9.5)	78 (12.6)
UNC	94 (4.6)	73 (5.6)	6 (5.7)	15 (2.4)
MT. SINAI	19 (0.9)	12 (0.9)	0 (0)	7 (1.1)
Age of ADI completion (month) (mean ± SD)	102 ± 68	103 ± 68	88 ± 59	101 ± 69
SOC (mean ± SD)	22.0 ± 5.6	22.1 ± 5.6	21.8 ± 5.8	21.9 ± 5.6
BEH (mean ± SD)	6.2 ± 2.5	6.3 ± 2.5	5.9 ± 2.4	6.1 ± 2.6

Table II lists the least squares means of the SOC and BEH scores by the four groups of affected subjects: male ASDs from the male-only families, male ASDs from the female-containing families, female ASDs from the female-only families and female ASDs from the female-containing families after the adjustment for covariate effects. The SOC scores were not significantly different across the four groups (P = 0.66) while the BEH scores were significantly different across the four groups (P < 0.0001). The order of the BEH scores from the smallest to the largest was: female from FC → female from FO → male from MO → male from FC. In general, the female ASDs had lower BEH scores than the male ASDs (overall female vs. male P < 0.0001). The mean (and SE) BEH values are demonstrated in Figure 1 according to the type of affected individual.

Table II. The SOC and BEH Scores From the Four Sex Groups

Family type (number)	Sex	Sample size	SOCa (LS mean ± SE)	BEHa (LS mean ± SE)
a Least squares means and standard errors after the adjustment for covariates, AGP site, age of ADI-R completion and verbal/non-verbal status.
Female-only (51)	Female	105	23.30 ± 0.55	5.77 ± 0.27
Female-containing (288)	Female	301	23.50 ± 0.33	5.53 ± 0.16
	Male	320	23.10 ± 0.32	6.35 ± 0.16
Male-only (631)	Male	1,302	23.14 ± 0.22	5.97 ± 0.11
Total		2,028	P = 0.66	P < 0.0001

Figure 1. The BEH scores for the four gender groups least squares means ±95% confidence interval.

In a post-hoc analysis, we looked at the individual items making up the BEH domain and noted that it was largely items measuring the higher order “insistence on sameness” factor identified in several previous reports [Turner, 1999; Georgiades et al., 2007; Lam et al., 2008; Smith et al., 2009], not the lower order “sensory motor” behaviors, that showed these overall sex differences in mean scores. For example, females had lower scores on unusual preoccupations (P < 0.001), circumscribed interests (P = 0.002), repetitive use of objects or interest in parts of objects (P = 0.03), and the “encompassing preoccupation or circumscribed pattern of interest” subdomain total score (P < .001) (which is the sum of items “unusual preoccupation” and “circumscribed interests”).

Table III lists the pair-wise comparisons of the BEH scores from the four groups of affected subjects. The largest BEH score difference was observed between the female and male ASDs from the FC families (difference in LS means = 0.82 and P < 0.0001). The key analysis focuses on the male siblings from the two types of families; male only (MO) and female containing (FC). The BEH scores from the male ASDs of the FC families was also significantly higher than the scores from the male ASDs of the MO families (difference in LS means = 0.38; P = 0.013). Virtually, the same items from the BEH domain that differentiated males and females in the earlier analysis also appeared to differentiate these two groups of males; unusual preoccupations (P = 0.01), circumscribed interests (P = 0.08), and the subdomain total score of these two items (P = 0.001).

Table III. Comparison of the BEH Scores From the Four Groups of Affected Subjects

	Female from FO	Female from FC	Male from FC	Male from MO
The differences in least squares mean ± standard error are in the lower triangle and the t-test, P values are in the upper triangle. The differences in least squares means were derived after the adjustment for covariates, AGP site, age of ADI-R completion and verbal/non-verbal status.
Female from FO	—	0.40	0.046	0.46
Female from FC	0.24 ± 0.29	—	<0.0001	0.0047
Male from FC	0.58 ± 0.29	0.82 ± 0.17	—	0.013
Male from MO	0.20 ± 0.27	0.44 ± 0.16	0.38 ± 0.15	—

CONCLUSIONS

We were able to confirm our first hypothesis that females with ASD are less severely affected than males with ASD on the repetitive stereotyped dimension of behavior. This differential severity by sex did not extend to measures of social reciprocity. The influence of sex on domains of ASD symptomatology has been reported previously but the results have not been consistent [see Lord et al., 1982; Carter et al., 2007; Holtmann et al., 2007; Hartley and Sikora, 2009]. This may be due to several factors. Those studies had small sample sizes and the differences reported in our study are quite small (Table III) and would require a large sample size to ensure sufficient power. Alternatively, studies that only sampled individuals with autism [and indeed narrower definitions of autism than are currently used, see Lord et al., 1982], presumably sampled a truncated part of the distribution of RSB's from a population of subjects with the broader phenotype of ASD. We sampled individuals with ASD according to the ADI-R but excluded those with even milder phenotypes to ensure misclassification was minimal. Both McLennan et al. [1993] and Carter et al. [2007] did demonstrate a trend for the BEH scores in their samples to be higher in the males than in the females though these differences were not statistically significant. Due to the expanding of diagnostic criteria, we may have sampled more higher functioning females with ASD compared to earlier studies of girls (which presumably largely all came from singleton families) perhaps making the detection of sex differences in this sub-phenotype easier. No sex differences existed on measures of social reciprocity in our analyses suggesting that the mechanism underlying sex differences in ASD is either restricted to the BEH domain or else there are measurement effects or other covariates that influence social reciprocity and so the familial effect is obscured. The results in this paper are also somewhat different than our earlier paper on sex differences in IQ which showed (in a sub-sample of the AGP from the Canadian site that was part of this study) that in families with a single affected individual, females are more severely affected than boys, whereas no such differences exist in multiple incidence families [Banach et al., 2009]. Putting these findings together adds to the growing evidence of the fractionation of the ASD phenotype into social-communication, and BEH [Happe and Ronald, 2008; Mandy and Skuse, 2008] and possibly IQ [Banach et al., 2009] with possibly different genotype-phenotype relationships. The findings also provide evidence that perhaps the factors that protect females from developing ASD in the first place may also protect them against high scores on the BEH domain of the ADI-R.

Our second hypothesis (that the affected male relatives of females with ASD will be more severely affected than the affected male relatives of male cases) was also confirmed. Insofar as the sex of one relative affects the quantitative trait of another relative, we are able to support the idea that the differences in severity in BEH between the sexes is under familial (presumably genetic) control. Secondary analyses showed that these differences in the BEH domain were largely restricted to items measuring insistence on sameness/ritualistic types of repetitive stereotyped behaviors not the sensory motor kinds. This adds to the growing evidence that the higher and lower order types of repetitive stereotyped behaviors are different and may arise from different genetic mechanisms [Georgiades et al., 2007; Lam et al., 2008; Smith et al., 2009].

We are not aware of any environmental factors that could explain the results reported here. We have derived these predictions based on a model of genetic heterogeneity and one of its subsets; the polygenic multifactorial model of sex-specific thresholds. In either case, the data fit the hypothesis which rest on the assumption that differences in severity of a sub-phenotype associated with sex are based on differences in genetic liability. This is not an unreasonable assumption and is one often offered as an explanation by several investigators [Tsai and Beisler, 1983; Ritvo et al., 1989; Volkmar et al., 1993].

Several potential limitations of this study deserve mention. First, the affected subjects that were included in this study are a heterogeneous group which could be an issue if the sex-specific polygenic model was only applicable to a certain ASD subtype. However, most of the ASD cases in the AGP meet strict ADI-R and ADOS criteria for autism so there is little power to test this hypothesis. The different sites in the AGP used slightly different inclusion and exclusion criteria and ascertainment strategies and that too may have lead to site-specific differences in spite of our attempts to control for site. It is perhaps worth pointing out that the impact of sex on BEH scores in affected relatives is stronger without site as a covariate. It is unfortunate that the sites all used different IQ tests perhaps making it impossible to use IQ as a covariate on a continuous scale. This will have reduced power to detect IQ as a confounding variable. Third, the result that these findings pertain to BEH and not to SOC may be the result of differences in measurement. The SOC domain of the ADI may have more measurement error associated with it or since the inclusion criteria for the sample was that they met a cut-off for the SOC domain but not the BEH domain, ceiling effects in the SOC domain may make it difficult to find differences between groups. However, the distribution of scores was symmetrical and this appeared to have no effect on the results. Finally, it is still possible that we did not have the power to detect differences on the SOC trait compared to the BEH trait since the standard deviations of the former are greater than those of the latter. It is important to remember that these results only apply to affected individuals from multiply affected families. No generalization to singletons is possible as the relationship between sex and IQ is different in these two types of families and suggests that differences in other sub-phenotypes may exist as well.

We conclude that sex differences exist with respect to differences in RSB between males and females with ASD in multiple incidence families and this can be partly explained on the basis of differences in genetic liability between males and females. What those factors are is an area worthy of further exploration as they would shed light on the genetic architecture of the disorder and the variables that contribute to the enormous differences in severity seen among children with ASD. As we search for those genetic and genomic variants associated with the disorder, not only is it important to be cognizant of the phenotype being addressed, and the sex of the proband, we can now add it is important to take into account the sex of another individual in the pedigree with ASD, if such a person exists. Of course, it still not clear why females from families with ASD have lower BEH scores than males with ASD. Possibly the genetic, or perhaps epigenetic, factors that protect girls from developing the disorder in the first place may also reduce BEH scores among those who are affected. If that is true, identifying these protective factors should be a priority in future studies of the etiology of the disorder.

Acknowledgements

The authors wish to acknowledge the contributions of the Autism Genome Project who collected the samples and made many useful comments during the composition of this paper. The ideas contained in this paper were first discussed with Drs. Neil Risch and Kathleen Merikangas. We also gratefully acknowledge the contributions of the families who participated in this study.

Support for the AGP when this work was carried out came from: Autism Speaks (USA); Genome Canada (Canada); Canadian Institutes of Health Research; the Health Research Board (HRB; Ireland); the Hilibrand Foundation (USA) and the Medical Research Council (MRC; UK).

Drs. Goldberg, Zwaigenbaum and Szatmari were supported by Fellowship awards from the Ontario Mental Health Foundation and operating grants from the Canadian Institutes of Health Research.